Paper deep dive
Evidential Neural Radiance Fields
Ruxiao Duan, Alex Wong
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 88%
Last extracted: 7/20/2026, 8:18:12 AM
Summary
The paper introduces Evidential Neural Radiance Fields (Evidential NeRF), a probabilistic framework that quantifies both aleatoric and epistemic uncertainties in 3D scene reconstruction via a single forward pass. It addresses limitations of existing NeRF uncertainty methods by using an evidential deep learning approach to separate data variability from model ignorance, achieving state-of-the-art performance on standardized benchmarks.
Entities (8)
Relation Signals (7)
Evidential Neural Radiance Fields → quantifies → Aleatoric uncertainty
confidence 95% · enabling direct quantification of both aleatoric and epistemic uncertainties from a single forward pass
Evidential Neural Radiance Fields → quantifies → Epistemic Uncertainty
confidence 95% · enabling direct quantification of both aleatoric and epistemic uncertainties from a single forward pass
Evidential Neural Radiance Fields → usesdistribution → Normal-Inverse-Gamma Distribution
confidence 90% · model (μ,σ2) by a normal-inverse-gamma (NIG) distribution
Evidential Neural Radiance Fields → outperforms → Normal NeRF
confidence 88% · our approach demonstrates state-of-the-art scene reconstruction fidelity and uncertainty estimation quality
Evidential Neural Radiance Fields → basedon → Evidential Deep Learning
confidence 85% · we adapt evidential reasoning to the volumetric structure of NeRFs
Evidential Neural Radiance Fields → evaluatedon → LF Dataset
confidence 80% · We compare multiple uncertainty quantification methods on three standardized benchmarks... LF
Evidential Neural Radiance Fields → evaluatedon → LLFF Dataset
confidence 80% · We compare multiple uncertainty quantification methods on three standardized benchmarks... LLFF
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Understanding sources of uncertainty is fundamental to trustworthy three-dimensional scene modeling. While recent advances in neural radiance fields (NeRFs) achieve impressive accuracy in scene reconstruction and novel view synthesis, the lack of uncertainty estimation significantly limits their deployment in safety-critical settings. Existing uncertainty quantification methods for NeRFs fail to separately capture both aleatoric and epistemic uncertainties. Among those that do quantify one or the other, many of them either compromise rendering quality or incur significant computational overhead to obtain uncertainty estimates. To address these issues, we introduce Evidential Neural Radiance Fields, a probabilistic approach that seamlessly integrates with the NeRF rendering process, enabling direct quantification of both aleatoric and epistemic uncertainties from a single forward pass. We compare multiple uncertainty quantification methods on three standardized benchmarks, where our approach demonstrates state-of-the-art scene reconstruction fidelity and uncertainty estimation quality. Code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2602.23574v2
- Canonical: https://arxiv.org/abs/2602.23574v2
Trouble viewing inline? Open PDF directly →
Full Text
80,546 characters extracted from source content.
Expand or collapse full text
Evidential Neural Radiance Fields Ruxiao Duan Yale University ruxiao.duan@yale.edu Alex Wong Yale University alex.wong@yale.edu Abstract Understanding sources of uncertainty is fundamental to trustworthy three-dimensional scene modeling. While recent advances in neural radiance fields (NeRFs) achieve impressive accuracy in scene reconstruction and novel view synthesis, the lack of uncertainty estimation significantly limits their deployment in safety-critical settings. Existing uncertainty quantification methods for NeRFs fail to separately capture both aleatoric and epistemic uncertainties. Among those that do quantify one or the other, many of them either compromise rendering quality or incur significant computational overhead to obtain uncertainty estimates. To address these issues, we introduce Evidential Neural Radiance Fields, a probabilistic approach that seamlessly integrates with the NeRF rendering process, enabling direct quantification of both aleatoric and epistemic uncertainties from a single forward pass. We compare multiple uncertainty quantification methods on three standardized benchmarks, where our approach demonstrates state-of-the-art scene reconstruction fidelity and uncertainty estimation quality. Code is available at https://github.com/KerryDRX/EvidentialNeRF. Figure 1: Scene reconstruction along with its aleatoric and epistemic uncertainties rendered by Evidential NeRF from a scene in the wild. AU arises from intrinsic data variability in the training images, including radiance variation (e.g., sky regions due to changing illumination), high-frequency regions (e.g., object edges), and the presence of transient objects (e.g., pedestrians). EU indicates the model’s lack of knowledge, prominently appearing in occluded regions where training supervision is insufficient (e.g., the trees occluded by the gate). 1 Introduction Predictive uncertainty of deep neural networks originates from two different sources, data and model, and the two types of uncertainty are respectively referred to as aleatoric uncertainty (AU) and epistemic uncertainty (EU) [18, 14]. Aleatoric (data) uncertainty stems from intrinsic randomness of the data generation process, while epistemic (model) uncertainty arises from model’s lack of knowledge [9, 12]. Understanding both types of predictive uncertainty is essential for building models that are not only accurate but also reliable and explainable under various conditions. Neural radiance fields (NeRFs) [26] have achieved remarkable performance in three-dimensional (3D) scene reconstruction and novel view synthesis, yet their incapability of quantifying predictive uncertainty poses significant challenges for their broader adoption in safety-critical domains such as autonomous driving [11], medical imaging [45], and robotics [44], where precise and prompt quantification of predictive uncertainty remains crucial. We consider an essential but overlooked research question: How should a NeRF report what it does not know (epistemic uncertainty) versus what the data cannot resolve (aleatoric uncertainty), without sacrificing rendering fidelity or speed? Existing uncertainty quantification (UQ) methods for NeRFs generally fall into three categories: closed-form likelihood models, which are unable to capture epistemic uncertainty [24, 30, 36, 35], Bayesian methods, which typically necessitate sampling during inference [7, 39, 37], and ensemble approaches, which require training multiple models [22, 41]. None of these methods quantifies both types of uncertainty with a single forward pass, and many of them have to sacrifice prediction accuracy to obtain uncertainty estimates. Among the likelihood models is the classical Gaussian approach [18] that represents a prediction as the mean and uncertainty as the variance of a normal distribution. This method has been widely adopted in not only NeRFs [24, 30, 31] but also other radiance field frameworks such as Gaussian Splatting [19, 21] and DVGO [40, 38]. While computationally efficient, this paradigm inherently models only aleatoric uncertainty while failing to account for epistemic uncertainty. To address these limitations, we propose Evidential Neural Radiance Fields (Evidential NeRFs), a probabilistic framework to separately quantify both aleatoric and epistemic uncertainties of NeRFs through a single forward pass (Figure 1). Our method extends the normal probabilistic formulation of radiance modeling by treating the predicted mean and variance of pixel radiance themselves as random variables governed by a higher-order evidential distribution, yielding closed-form predictive uncertainties through the NeRF rendering process. Unlike prior evidential deep learning methods [34, 1], which regress evidential distribution parameters, we adapt evidential reasoning to the volumetric structure of NeRFs, enabling them to predict aleatoric and epistemic uncertainties directly instead of having to reformulate them from evidential parameters. Beyond the technical limitations of existing approaches, which we address with Evidential NeRF, we also observed that current uncertainty quantification benchmarks lack standardization. Methods are often evaluated using different architectures, data splits, and training setups, making direct comparisons difficult. To mitigate these confounding factors, we establish a new benchmark to isolate and evaluate the UQ methods themselves. Under this standardized evaluation, our method consistently ranks within the top three across all image reconstruction and uncertainty quantification metrics, if not the first. The remaining top-performing methods are mostly ensemble-based, which incur substantial computational cost and are unsuitable for real-time applications. In contrast, our approach is the second-fastest overall, being only 0.04 FPS slower than the fastest method while delivering significantly better performance across all metrics. Our contributions can be summarized as follows. • We present a probabilistic framework for neural radiance fields to separately quantify both aleatoric and epistemic uncertainties in 3D scene reconstructions. • We provide detailed mathematical derivations showing how aleatoric and epistemic uncertainties can be propagated from points to pixels under proper independence assumptions, enabling seamless integration of evidential deep learning with the volumetric rendering paradigm. • We establish a standardized benchmark for faithful comparison of NeRF uncertainty quantification methods. Figure 2: Evolution of NeRF pipeline across three levels of probabilistic formulations. N points sampled along the camera ray give N pairs of spatial location and viewing direction, which are passed to the NeRF model for prediction. Level 1: Vanilla NeRF predicts only point density and color, resulting in a deterministic pixel color without any uncertainty estimate. Level 2: Normal NeRF assumes the point and pixel colors follow normal distributions, quantifying aleatoric uncertainty of rendered color. Level 3: Evidential NeRF assumes the pixel color has random mean and variance following an evidential distribution, quantifying both aleatoric and epistemic uncertainties. 2 Related Work Neural radiance fields. NeRFs [26] represent a 3D scene as a plenoptic function which can be learned from a sparse set of input views. Realized as a deep neural network, the plenoptic function maps spatial locations and viewing directions to their volume densities and radiance. NeRFs produce high-fidelity reconstructions and photorealistic novel view renderings. Numerous variants in architectures, loss functions, and sampling strategies have been proposed to further improve the model’s training efficiency, inference speed, and rendering quality [29, 15, 6, 5, 43, 27, 2, 3, 8, 13, 46, 4]. NeRF uncertainty quantification. UQ methods for NeRFs generally follow three schools of thought. First, closed-form likelihood models. NeRF-W [24] and ActiveNeRF [30] follow the same probabilistic framework in [18] to model ray colors by normal distributions. MixNeRF [36] and FlipNeRF [35], though not proposed as UQ methods, model ray colors by mixtures of Laplace distributions to represent the multimodal nature of radiance. These approaches employ closed-form probabilistic distributions to model pixel colors, enabling efficient training and uncertainty inference, but overlooking epistemic uncertainty, due to the deterministic nature of the distributions employed. Second, Bayesian methods. Monte Carlo dropout [7] estimates epistemic uncertainty by interpreting dropout as variational inference and performing multiple stochastic forward passes at test time. S-NeRF [39] and CF-NeRF [37] learn a probability distribution over all the possible radiance fields and approximate the posterior distribution with stochastic variational inference [20] and conditional normalizing flows [32]. While these approaches provide epistemic uncertainty estimates, they typically incur significant computational overhead due to sampling at inference. Third, ensemble methods. Deep ensembles [22] estimate epistemic uncertainty by training multiple models independently and measuring their prediction variance during inference. Density-aware NeRF Ensembles (DANE) [41] further incorporates a density-based epistemic term that captures uncertainty in scene geometry and appearance, which is added to the uncertainty estimated by standard ensembles. Ensemble-based methods impose the most significant computation and memory overheads, as each model must be trained, stored, and separately evaluated during inference. Evidential deep learning. Based on the theory of subjective logic [17], evidential deep learning (EDL) [34] is first introduced for classification uncertainty quantification based on a single network and later extended to regression [1]. EDL considers model learning as a process of evidence collection, in which each training sample adds support to a higher-order evidential distribution, which is a Dirichlet distribution for classification and a normal-inverse-gamma (NIG) distribution for regression [34, 1]. The probability distribution from which predictions are drawn is assumed to have random parameters sampled from the evidential distribution, and the aleatoric and epistemic predictive uncertainties can be directly obtained by a single forward propagation. Though effective in standard regression, EDL is not naively compatible with NeRF learning paradigm due to its hierarchical volumetric formulation: As NeRFs receive supervision after volumetric rendering, evidential parameters tied to point-level predictions cannot be learned directly from pixel-level observations. Concurrent with our work, ENeRF [47] proposes evidential modeling in NeRFs by adopting NIG distribution at the point level and approximating pixel-level parameters via the mixture of NIGs [23]. 3 Preliminaries Neural radiance fields. NeRF learns a continuous scene representation function f by a multilayer perceptron that maps a pair of spatial location (referred to as point in this paper) i∈ℝ3 x_i ^3 and unit viewing direction ∈ℝ3 d ^3 to its corresponding view-independent volume density ρi>0 _i>0 and directional emitted radiance i∈[0,1]3 c_i∈[0,1]^3 in RGB color. For notational simplicity, consider a scalar color channel ci∈[0,1]c_i∈[0,1]. A NeRF model predicts ci,ρi=f(i,).c_i, _i=f( x_i, d). (1) The pixel color of a ray (t)=+t r(t)= o+t d defined by a camera center o and a viewing direction d is derived by first sampling N points (ti)i=1N\ r(t_i)\_i=1^N along the ray and then computing the weighted sum of point colors via the discretized volumetric rendering equation c=∑i=1Nwicic= _i=1^Nw_ic_i, where wi=exp(−∑j=1i−1ρjδj)(1−exp(−ρiδi))w_i= (- _j=1^i-1 _j _j ) (1- (- _i _i) ) is the weight of the i-th point and δi=ti+1−ti _i=t_i+1-t_i is the distance between two adjacent points. Normal neural radiance fields. Based on the probabilistic framework in [18], a point color cic_i can be modeled by a normal distribution as ci|μi,σi2∼(μi,σi2),c_i\ |\ _i, _i^2 ( _i, _i^2), (2) with mean μi∈[0,1] _i∈[0,1] and variance σi2>0 _i^2>0 predicted by (μi,σi2),ρi=f(i,).( _i, _i^2), _i=f( x_i, d). (3) As the pixel color c is the weighted sum of point colors cic_i’s, c also follows a normal distribution assuming independence of points: c|μ,σ2∼(μ,σ2),c\ |\ μ,σ^2 (μ,σ^2), (4) where μ≔∑i=1Nwiμiμ _i=1^Nw_i _i is the mean (prediction) and σ2≔∑i=1Nwi2σi2σ^2 _i=1^Nw_i^2 _i^2 is the variance (uncertainty) of pixel color. While adjacent points in practice exhibit statistical dependence in radiance, the independence assumption is a common and effective simplification to enable tractable aggregation of point uncertainties into the pixel level. Though efficient, this approach only captures aleatoric uncertainty. 4 Method Based on Equations 2 and 4, we establish an evidential probabilistic framework for radiance modeling by taking a step further to assume random conditional mean and variance. Different from conventional EDL approaches which train the model to predict evidential parameters, we instead let the model predict aleatoric and epistemic uncertainties directly (Section 4.1), and then propagate the uncertainties from points to pixels compatible with the volumetric rendering paradigm of NeRFs (Section 4.2). Then the evidential distribution can be reformulated from the rendered color and uncertainties (Section 4.3) and learned at the pixel level where supervision is available (Section 4.4). 4.1 Point-level Probabilistic Radiance Modeling Building upon Equation 2, we further treat the conditional mean and variance of point radiance as random variables instead of point estimates, i.e., (μi,σi2)∼πi( _i, _i^2) _i where πi _i is some probability distribution. Under this formulation, the predictive mean, total, aleatoric, and epistemic uncertainties of the point color cic_i can be respectively formulated as c¯i c_i ≔[ci]=[[ci|μi,σi2]]=[μi], [c_i]=E[E[c_i| _i, _i^2]]=E[ _i], (5) Ui U_i ≔Var[ci]=Uialea+Uiepis, [c_i]=U_i^alea+U_i^epis, (6) Uialea U_i^alea ≔[Var[ci|μi,σi2]]=[σi2], [Var[c_i| _i, _i^2]]=E[ _i^2], (7) Uiepis U_i^epis ≔Var[[ci|μi,σi2]]=Var[μi]. [E[c_i| _i, _i^2]]=Var[ _i]. (8) Note: The classical assumption that μi _i and σi2 _i^2 are point estimates corresponds to the special case where πi _i is a degenerate distribution with a Dirac delta density function concentrated at fixed values of μ^i μ_i and σ^i2 σ_i^2 and the uncertainties above are thereby reduced to Uialea=[σi2]=σ^i2U_i^alea=E[ _i^2]= σ_i^2 and Uiepis=Var[μi]=0U_i^epis=Var[ _i]=0, in which case only aleatoric uncertainty can be captured. An Evidential NeRF model predicts (c¯i,Uialea,Uiepis,α~i),ρi=f(i,),( c_i,U_i^alea,U_i^epis, α_i), _i=f( x_i, d), (9) where c¯i∈[0,1] c_i∈[0,1] is the point’s mean color, Uialea>0U_i^alea>0 and Uiepis>0U_i^epis>0 are aleatoric and epistemic uncertainties, and α~i>0 α_i>0 is a positive shape score which will later be used to derive a pixel-level shape parameter α in Section 4.3. We apply sigmoid activation to predict c¯i c_i as in original NeRF and softplus activation for all other parameters to enforce range restrictions. The model architecture remains mostly unchanged, except for the three additional output neurons in the last layer to predict (Uialea,Uiepis,α~i)(U_i^alea,U_i^epis, α_i). 4.2 Propagation from Points to Pixels Let the set of conditional means and variances of all point colors along a ray be ≔(μi,σi2)i=1N θ \( _i, _i^2)\_i=1^N, which fully specifies the conditional distribution of the ray color c. Under this setting, the pixel’s predictive mean color, total, aleatoric, and epistemic uncertainties are defined as c¯ c ≔[c], [c], U U ≔Var[c], [c], (10) Ualea U^alea ≔[Var[c|]], [Var[c| θ]], Uepis U^epis ≔Var[[c|]]. [E[c| θ]]. (11) With the point-level predictions from Equation 9 and the independence assumption of points, the pixel color and uncertainties can be derived as c¯ c =∑i=1Nwic¯i, = _i=1^Nw_i c_i, U U =∑i=1Nwi2Ui, = _i=1^Nw_i^2U_i, (12) Ualea U^alea =∑i=1Nwi2Uialea, = _i=1^Nw_i^2U_i^alea, Uepis U^epis =∑i=1Nwi2Uiepis. = _i=1^Nw_i^2U_i^epis. (13) In other words, pixel-level aleatoric and epistemic uncertainties can be obtained as weighted sums of their point-level counterparts, where the weights are equal to the squared weights for color volumetric rendering. Detailed proofs are provided in Section 8. 4.3 Pixel-level Probabilistic Radiance Modeling Now we introduce the probabilistic model for pixel colors. Recall that μ=∑i=1Nwiμiμ= _i=1^Nw_i _i and σ2=∑i=1Nwi2σi2σ^2= _i=1^Nw_i^2 _i^2 represent the conditional mean and variance of the pixel color c in Equation 4. To quantify aleatoric and epistemic uncertainties of a pixel color in closed form, we model (μ,σ2)(μ,σ^2) by a normal-inverse-gamma (NIG) distribution μ,σ2∼NIG(γ,ν,α,β),μ,σ^2 (γ,ν,α,β), (14) or equivalently, μ|σ2∼(γ,σ2/ν)μ\ |\ σ^2 (γ,σ^2/ν) and σ2∼Γ−1(α,β)σ^2 ^-1(α,β), where γ∈[0,1]γ∈[0,1], ν>0ν>0, and Γ−1(α,β) ^-1(α,β) denotes an inverse-gamma distribution with shape α>1α>1 and scale β>0β>0. Consider each pixel color c as being generated through a hierarchical sampling process: Given the higher-order evidential distribution NIG(γ,ν,α,β)NIG(γ,ν,α,β), drawing a sample (μ,σ2)(μ,σ^2) from it yields an instance of lower-order normal distribution (μ,σ2)N(μ,σ^2), from which the pixel color c is sampled. With this formulation, the pixel color’s predictive mean, total, aleatoric, and epistemic uncertainties can be directly expressed in terms of the NIG parameters: c¯ c =[c]=[[c|μ,σ2]]=[μ]=γ, =E[c]=E[E[c|μ,σ^2]]=E[μ]=γ, (15) U U =Var[c]=Ualea+Uepis, =Var[c]=U^alea+U^epis, (16) Ualea U^alea =[Var[c|μ,σ2]]=[σ2]=βα−1, =E[Var[c|μ,σ^2]]=E[σ^2]= βα-1, (17) Uepis U^epis =Var[[c|μ,σ2]]=Var[μ]=β(α−1)ν. =Var[E[c|μ,σ^2]]=Var[μ]= β(α-1)ν. (18) Dataset Method PSNR↑ SSIM↑ LPIPS↓ NLL↓ AUSE RMSE↓ AUSE MAE↓ LF [48] Baseline 28.5538 0.9172 0.0465 Dropout [7] 28.1137 0.9061 0.0569 3.6670 0.0125 0.0049 Normal [18] 28.0064 0.9165 0.0531 0.4425 0.0090 tabthird0.0029 MoL [36] 28.2200 0.9095 0.0672 tabfirst-2.5393 tabthird0.0084 tabthird0.0029 Ensembles [22] tabsecond29.3779 tabsecond0.9308 tabsecond0.0411 0.3245 tabfirst0.0070 tabsecond0.0026 DANE [41] tabsecond29.3779 tabsecond0.9308 tabsecond0.0411 tabthird-0.4317 0.0101 0.0039 Evidential tabfirst29.9679 tabfirst0.9345 tabfirst0.0359 tabsecond-2.4491 tabfirst0.0070 tabfirst0.0025 LLFF [25] Baseline 17.5281 0.4720 0.4119 Dropout [7] 17.1423 0.4441 0.4468 90.6773 0.0742 0.0426 Normal [18] 16.5500 0.4135 0.4732 55.3580 0.0674 0.0405 MoL [36] 16.4218 0.4237 0.4842 tabsecond2.2470 0.0814 0.0384 Ensembles [22] tabfirst17.9181 tabfirst0.5109 tabsecond0.3932 11.1658 tabfirst0.0513 tabfirst0.0253 DANE [41] tabfirst17.9181 tabfirst0.5109 tabsecond0.3932 tabthird9.7273 tabsecond0.0521 tabsecond0.0260 Evidential tabthird17.8793 tabthird0.5068 tabfirst0.3751 tabfirst0.6765 tabthird0.0578 tabthird0.0295 RobustNeRF [33] Baseline 25.2205 0.8296 0.1577 Dropout [7] 24.7693 0.8118 0.1745 22.8799 0.0284 0.0160 Normal [18] 25.2987 0.8522 tabsecond0.1311 10.4873 tabthird0.0250 tabthird0.0151 MoL [36] 23.7874 0.7921 0.2049 tabfirst-1.3947 0.0304 0.0176 Ensembles [22] tabsecond26.1953 tabsecond0.8562 tabthird0.1438 4.6309 tabfirst0.0164 tabfirst0.0098 DANE [41] tabsecond26.1953 tabsecond0.8562 tabthird0.1438 tabthird4.1092 0.0283 0.0154 Evidential tabfirst26.2292 tabfirst0.8641 tabfirst0.1112 tabsecond-1.2702 tabsecond0.0221 tabsecond0.0138 Table 1: Quantitative results of scene reconstruction and uncertainty quantification on three datasets, averaged over three independent runs. Colored cells denote the first, second, and third best results. See per-scene statistics with standard deviations in supplementary Section 10. Since the Evidential NeRF model directly predicts uncertainties and shape scores instead of NIG parameters, the NIG parameters need to be reformulated as γ γ =c¯, = c, ν ν =UaleaUepis, = U^aleaU^epis, (19) α α =1+∑i=1Nw~α~i, =1+ _i=1^N w α_i, β β =Ualea(α−1), =U^alea(α-1), (20) where c¯ c, UaleaU^alea, and UepisU^epis are obtained from the point-to-pixel propagation process in Equations 12 and 13, α~i α_i is the point’s shape score, and w~i≔wi/∑j=1Nwj w_i w_i/ _j=1^Nw_j is the normalized weight that determines how much a point’s shape score contributes to the NIG shape parameter α of the pixel. Figure 3: Qualitative comparison on two example scenes, with image reconstructions, error maps, and uncertainty maps. Histogram equalization is conducted on the error maps to highlight the error regions. Our method’s uncertainty is total uncertainty. In general, our method achieves better reconstruction accuracy and produces uncertainty maps that are more consistent with prediction errors. 4.4 Learning Based on Equations 4 and 14, a pixel color c marginally follows a Student’s t distribution c∼t(γ,β(ν+1)αν,2α),c t (γ, β(ν+1)αν,2α ), (21) where t(μt,σt2,νt)t( _t, _t^2, _t) denotes a Student’s t distribution with location μt _t, scale σt _t, and degrees of freedom νt _t. The model is trained via maximum likelihood estimation by minimizing the negative log-likelihood (NLL) of the ground truth ℒnll _nll =−logp(cgt|γ,ν,α,β) =- p(c^gt|γ,ν,α,β) (22) =12logπν−αlogΩ+logΓ(α)Γ(α+12) = 12 πν-α + (α) (α+ 12 ) +(α+12)log((cgt−γ)2ν+Ω), + (α+ 12 ) ((c^gt-γ)^2ν+ ), (23) where cgt∈[0,1]c^gt∈[0,1] is the ground truth pixel color, Ω=2β(ν+1) =2β(ν+1), and Γ represents the Gamma function. See derivations in Section 9. Different values of ν and β may result in the same scale of the t distribution so long as the ratio β(ν+1)/νβ(ν+1)/ν is fixed. To resolve this ambiguity and also to suppress excessive evidence assigned to inaccurate predictions, a regularizer [1] is introduced to the loss function as ℒreg=|cgt−γ|(2ν+α),L_reg=|c^gt-γ|(2ν+α), (24) where |cgt−γ||c^gt-γ| is the absolute error of prediction and 2ν+α2ν+α is the count of virtual observations representing the evidence [1]. Therefore, the total loss is written as ℒ=ℒnll+λregℒreg,L=L_nll+ _regL_reg, (25) where λreg>0 _reg>0 is the regularization coefficient. Lastly, we generalize our color channel assumption from a single channel to three for RGB color modeling. We assume that the three color channels have different means but the same uncertainty, due to the high correlation of variance in different color channels of the same pixel [24, 30]. Practically, the output dimension of color prediction head in Equation 9 is set to three, predicting the point’s mean RGB color as ¯i∈[0,1]3 c_i∈[0,1]^3, and the uncertainty-related parameters are broadcast to all color channels during training. Figure 4: Test uncertainties vs. training sample size on Android. 5 Experiments 5.1 Settings Datasets. We test the UQ methods on three 3D scene reconstruction datasets: Light Field (LF) [48], Local Light Field Fusion (LLFF) [25], and RobustNeRF [33]. For LF, we test on the four scenes following the setup in [10] with the same train-test split. For LLFF, we employ its eight scenes and adopt the train-test split of [29], training with only three input views to assess UQ performance under sparse-view conditions. For RobustNeRF, we evaluate on its four scenes following their protocol of training on cluttered images only and testing on clean ones. Metrics. We evaluate the method performance based on the accuracy of both the rendered images and the uncertainty estimates. For images, we report PSNR, SSIM, and LPIPS to reflect the image reconstruction quality. For uncertainties, we use negative log-likelihood (NLL) and area under sparsification error (AUSE) with respect to both RMSE and MAE, measuring the quality of uncertainty estimates in terms of distributional fit and error ranking capabilities. Baselines. We select UQ methods from each of the three categories discussed in Section 2 as baselines. For closed-form likelihood models, we include the normal distribution approach [18] adopted in [24] and [30] and mixture of Laplace distributions (MoL) employed in [36] and [35]. Among Bayesian methods, we choose the classical Monte Carlo dropout approach [7]. For ensemble methods, we consider naive deep ensembles [22] and density-aware NeRF ensembles (DANE) [41]. Mode Baseline Dropout Normal MoL Ens. DANE Ours Train/min.↓ 11.84 88.54 12.22 12.71 59.22 59.22 13.57 Infer/FPS↑ 4.88 0.09 4.71 4.42 0.96 0.96 4.67 Table 2: Average training time per 30k30k steps and inference FPS of baseline nerfacto and different UQ methods on an A6000 GPU. Implementation. Prior benchmarks on NeRF uncertainty quantification suffer from varying choices of data splits, model architectures, and training schemes. To isolate UQ method effects from engineering confounders, we establish a new standardized benchmark to focus comparison on the underlying UQ approaches themselves. First, for data split, we follow the aforementioned scheme for all models. Second, for architecture, we use nerfacto model in [42] to implement all methods for efficiency. For likelihood models, the only difference is the output layer size and the loss function. For the Bayesian method, Monte Carlo dropout is implemented with a dropout probability 0.20.2, trained once and sampled five times at inference. For ensembles, both naive ensembles and DANE are implemented by training five models independently. Third, in training, we adopt the default batch size, optimizer, and learning rate scheduler of nerfacto for all methods and train all models for the same number of iterations, chosen based on the convergence speed on each scene. For stability, we run each method independently three times and report the averaged metrics. 5.2 Results Quantitative results. The quantitative performance of the uncertainty quantification methods is detailed in Table 1 with per-scene statistics provided in Section 10. Figure 5: A case where AU dominates EU. The highly reflective surface of the display case in the foreground incurs specular reflections. AU arises due to the presence of data noise caused by the inconsistency of light across different training views. For scene reconstruction, compared to the baseline nerfacto model, likelihood and Bayesian methods (Normal, MoL, Dropout) struggle to preserve the image reconstruction accuracy, indicating that their predictive performance is compromised in exchange for uncertainty estimation. Ensemble-based methods provide faithful reconstructions in general, but they are extremely inefficient due to their substantial computational cost in both training and inference. In contrast, Evidential NeRF, while being efficient, consistently outperforms the baseline in image reconstruction, demonstrating that our uncertainties need not come at the expense of rendering fidelity. Remarkably, with only a single network trained and a single forward pass required at inference, the evidential approach achieves prediction accuracy comparable to, and in 77 out of 99 image reconstruction metrics even surpassing, the computationally expensive ensemble methods. For uncertainty quantification, in terms of NLL, MoL model achieves the best result overall likely due to its multimodal radiance distribution assumption. Compared to the normal distribution with fixed mean and variance, a normal distribution with random mean and variance increases test data likelihood by factors of 1.8×1011.8× 10^1, 5.6×10235.6× 10^23, and 1.3×1051.3× 10^5 on the three datasets, respectively. This tremendous improvement is because the fixed mean and variance assumption used by classical normal-based UQ only captures the aleatoric variations within the training data while overlooking epistemic uncertainty. The likelihood gain is less pronounced on LF because its test images are adjacent to the training views, thus the distribution shift is relatively insignificant. As for AUSE, ensemble-based methods achieve the strongest results in general, thanks to their ability to represent predictive variability through multiple models. Nevertheless, our evidential approach remains highly competitive, often ranking second only to ensembles, demonstrating that jointly modeling aleatoric and epistemic factors yields uncertainty estimates with stronger correlation with errors. Qualitative results. We qualitatively compare our Evidential NeRF against other methods in Figure 3 and Section 10. A primary goal of UQ is to produce uncertainty maps that accurately identify potential model failures. Previous methods, by neglecting either the aleatoric or epistemic component of predictive uncertainty, often fall short of accurately delimiting regions of predictive inaccuracy. In contrast, by considering both components, our method demonstrates superior alignment with the reconstruction error maps. This suggests that jointly modeling data noise and the model knowledge gap is essential for generating the most reliable uncertainty estimates. Uncertainty scaling with data. We investigate the effect of training sample size on the magnitude of aleatoric and epistemic uncertainties using an example scene from RobustNeRF in Figure 4. We reserve a held-out set of views for evaluation and incrementally select 1010 to 5050 images from the remaining data for training. As the training sample size increases, the test AU increases while the test EU decreases in general, indicating that the additional, potentially more cluttered observations introduce greater data variability, whereas the model’s lack of knowledge is gradually mitigated with more training data. Figure 6: A case where EU dominates AU. The model is trained only by 55 images from the front and asked to render the scene from all the viewing angles. EU arises due to the lack of knowledge during training on the views out of the training distribution. Aleatoric and epistemic factors in 3D scenes. Several scenarios where aleatoric and epistemic uncertainties emerge are respectively presented in Figures 1, 5 and 6. Figure 1 illustrates an in-the-wild scene with several aleatoric factors (varying illumination, high-frequency edge regions, transient objects) and an epistemic factor (partial occlusions). Figure 5 showcases a scenario with a reflective object, where aleatoric light variation is the dominant source of potential rendering error. Figure 6 demonstrates an out-of-distribution example where the model is asked to render the scene from angles that never appeared during training, showing that the lack of model knowledge about these unseen views is the primary cause of predictive failure. Computational efficiency. We report the total training time and inference frames per second (FPS) of baseline nerfacto and different UQ methods in Table 2. Compared to other UQ methods, in training, our evidential approach is only slightly slower than other two likelihood models while being significantly faster than the ensemble methods. As for inference, since our approach directly predicts uncertainty fields, it is more efficient than the methods that require computing uncertainty from the predicted parameters, such as MoL. Since nerfacto runs on Tiny CUDA Neural Networks [28], which does not support dropout, dropout is implemented without acceleration, making it naturally slower than other methods. Figure 7: Scene cleaning based on aleatoric uncertainty as a post-processing step for floater removal. Points with AU above a certain threshold have their density reduced to become more transparent. By reducing the threshold, more artifacts can be eliminated. 5.3 Applications We demonstrate two example applications using aleatoric and epistemic uncertainties respectively in Figures 7 and 8. Scene cleaning. Training NeRFs on unconstrained images often leads to inaccurate predictions due to radiance inconsistencies from uncontrolled lighting changes or transient objects, which are irreducible data noise explained by aleatoric uncertainty. As a robust indicator of geometric artifacts, AU-based scene cleaning can be applied as a post-processing procedure to eliminate noise in the renderings. Figure 7 presents an example where the erroneous floaters of the rendering can be progressively removed by gradually changing the AU threshold to suppress point densities. Active learning. As a proxy of lack of model knowledge, epistemic uncertainty is an ideal metric for active learning. Figure 8 illustrates a next-best-view planning experiment on an LLFF scene. Starting from an initial training set of 55 images, an Evidential NeRF is trained iteratively by 55 rounds. In each round, the model is trained by 55 epochs with all the current training images and then tested on a holdout test set, before choosing 55 additional images from the remaining data to add to the training pool. Two sampling strategies for data selection are compared: choosing the images with the most EU and choosing images randomly. The average and standard deviation of three independent runs of each selection scheme are reported. Active selection based on EU gives noticeably higher PSNR than random selection, indicating the samples with higher epistemic uncertainty are more informative for model learning. Figure 8: Mean and standard derivation of test PSNR of three runs on Horns scene, with two active sampling strategies: EU-based selection and random selection. The samples identified via epistemic uncertainty are more informative for model learning. 6 Conclusion Despite the efficacy of NeRFs in photorealistic 3D scene reconstruction, the absence of robust predictive uncertainty quantification significantly hinders their deployment in safety-critical domains. While some uncertainty quantification methods have been proposed for NeRFs, none of them provides a mechanism for quantifying both aleatoric and epistemic uncertainties of the scene reconstructions. In this paper, we introduce Evidential Neural Radiance Fields, a principled probabilistic framework that resolves this critical gap. Our approach seamlessly adapts evidential deep learning to the hierarchical nature of NeRFs through a point-to-pixel aleatoric and epistemic uncertainty propagation paradigm. Through extensive experiments on three standardized benchmarks, Evidential NeRF demonstrates its superiority in both image reconstruction fidelity and uncertainty estimation quality compared to other methods, while identifying the aleatoric and epistemic factors that contribute to predictive uncertainty within a variety of 3D scenes. Limitation. For tractable uncertainty propagation, Evidential NeRF assumes deterministic volumetric density, leaving spatial uncertainty not explicitly modeled. Extending the formulation to capture the uncertainty in scene geometries would enable more comprehensive uncertainty modeling. 7 Acknowledgment This work is supported by NSF 2112562 Athena AI Institute. References [1] A. Amini, W. Schwarting, A. Soleimany, and D. Rus (2020) Deep evidential regression. Advances in neural information processing systems 33, p. 14927–14937. Cited by: §1, §2, §4.4, §4.4. [2] J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan (2021) Mip-nerf: a multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, p. 5855–5864. Cited by: §2. [3] J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman (2022) Mip-nerf 360: unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 5470–5479. Cited by: §2. [4] J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman (2023) Zip-nerf: anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 19697–19705. Cited by: §2. [5] A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su (2022) Tensorf: tensorial radiance fields. In European conference on computer vision, p. 333–350. Cited by: §2. [6] Z. Chen, T. Funkhouser, P. Hedman, and A. Tagliasacchi (2023) Mobilenerf: exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 16569–16578. Cited by: §2. [7] Y. Gal and Z. Ghahramani (2016) Dropout as a bayesian approximation: representing model uncertainty in deep learning. In international conference on machine learning, p. 1050–1059. Cited by: §1, §2, Table 1, Table 1, Table 1, §5.1. [8] S. J. Garbin, M. Kowalski, M. Johnson, J. Shotton, and J. Valentin (2021) Fastnerf: high-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, p. 14346–14355. Cited by: §2. [9] J. Gawlikowski, C. R. N. Tassi, M. Ali, J. Lee, M. Humt, J. Feng, A. Kruspe, R. Triebel, P. Jung, R. Roscher, et al. (2023) A survey of uncertainty in deep neural networks. Artificial Intelligence Review 56 (Suppl 1), p. 1513–1589. Cited by: §1. [10] L. Goli, C. Reading, S. Sellán, A. Jacobson, and A. Tagliasacchi (2024) Bayes’ rays: uncertainty quantification for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20061–20070. Cited by: §5.1. [11] L. He, L. Li, W. Sun, Z. Han, Y. Liu, S. Zheng, J. Wang, and K. Li (2024) Neural radiance field in autonomous driving: a survey. arXiv preprint arXiv:2404.13816. Cited by: §1. [12] W. He, Z. Jiang, T. Xiao, Z. Xu, and Y. Li (2024) A survey on uncertainty quantification methods for deep learning. External Links: 2302.13425, Link Cited by: §1. [13] T. Hu, S. Liu, Y. Chen, T. Shen, and J. Jia (2022) Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 12902–12911. Cited by: §2. [14] E. Hüllermeier and W. Waegeman (2021) Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods. Machine learning 110 (3), p. 457–506. Cited by: §1. [15] A. Jain, M. Tancik, and P. Abbeel (2021) Putting nerf on a diet: semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 5885–5894. Cited by: §2. [16] Y. Jin, D. Mishkin, A. Mishchuk, J. Matas, P. Fua, K. M. Yi, and E. Trulls (2021) Image matching across wide baselines: from paper to practice. International Journal of Computer Vision 129 (2), p. 517–547. Cited by: §10. [17] A. Jsang (2018) Subjective logic: a formalism for reasoning under uncertainty. Springer Publishing Company, Incorporated. Cited by: §2. [18] A. Kendall and Y. Gal (2017) What uncertainties do we need in bayesian deep learning for computer vision?. Advances in neural information processing systems 30. Cited by: §1, §1, §2, §3, Table 1, Table 1, Table 1, §5.1. [19] B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis (2023) 3D gaussian splatting for real-time radiance field rendering.. ACM Trans. Graph. 42 (4), p. 139–1. Cited by: §1. [20] D. P. Kingma (2013) Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114. Cited by: §2. [21] J. Kulhanek, S. Peng, Z. Kukelova, M. Pollefeys, and T. Sattler (2024) Wildgaussians: 3d gaussian splatting in the wild. arXiv preprint arXiv:2407.08447. Cited by: §1. [22] B. Lakshminarayanan, A. Pritzel, and C. Blundell (2017) Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems 30. Cited by: §1, §2, Table 1, Table 1, Table 1, §5.1. [23] H. Ma, Z. Han, C. Zhang, H. Fu, J. T. Zhou, and Q. Hu (2021) Trustworthy multimodal regression with mixture of normal-inverse gamma distributions. Advances in Neural Information Processing Systems 34, p. 6881–6893. Cited by: §2. [24] R. Martin-Brualla, N. Radwan, M. S. Sajjadi, J. T. Barron, A. Dosovitskiy, and D. Duckworth (2021) Nerf in the wild: neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 7210–7219. Cited by: §1, §2, §4.4, §5.1. [25] B. Mildenhall, P. P. Srinivasan, R. Ortiz-Cayon, N. K. Kalantari, R. Ramamoorthi, R. Ng, and A. Kar (2019) Local light field fusion: practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG) 38 (4), p. 1–14. Cited by: Table 1, §5.1. [26] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng (2020) NeRF: representing scenes as neural radiance fields for view synthesis. In ECCV, Cited by: §1, §2. [27] T. Müller, A. Evans, C. Schied, and A. Keller (2022) Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG) 41 (4), p. 1–15. Cited by: §2. [28] tiny-cuda-n External Links: Link Cited by: §5.2. [29] M. Niemeyer, J. T. Barron, B. Mildenhall, M. S. Sajjadi, A. Geiger, and N. Radwan (2022) Regnerf: regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 5480–5490. Cited by: §2, §5.1. [30] X. Pan, Z. Lai, S. Song, and G. Huang (2022) Activenerf: learning where to see with uncertainty estimation. In European Conference on Computer Vision, p. 230–246. Cited by: §1, §2, §4.4, §5.1. [31] W. Ren, Z. Zhu, B. Sun, J. Chen, M. Pollefeys, and S. Peng (2024) Nerf on-the-go: exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8931–8940. Cited by: §1. [32] D. Rezende and S. Mohamed (2015) Variational inference with normalizing flows. In International conference on machine learning, p. 1530–1538. Cited by: §2. [33] S. Sabour, S. Vora, D. Duckworth, I. Krasin, D. J. Fleet, and A. Tagliasacchi (2023) Robustnerf: ignoring distractors with robust losses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 20626–20636. Cited by: Table 1, §5.1. [34] M. Sensoy, L. Kaplan, and M. Kandemir (2018) Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems 31. Cited by: §1, §2. [35] S. Seo, Y. Chang, and N. Kwak (2023) Flipnerf: flipped reflection rays for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 22883–22893. Cited by: §1, §2, §5.1. [36] S. Seo, D. Han, Y. Chang, and N. Kwak (2023) Mixnerf: modeling a ray with mixture density for novel view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20659–20668. Cited by: §1, §2, Table 1, Table 1, Table 1, §5.1. [37] J. Shen, A. Agudo, F. Moreno-Noguer, and A. Ruiz (2022) Conditional-flow nerf: accurate 3d modelling with reliable uncertainty quantification. In European Conference on Computer Vision, p. 540–557. Cited by: §1, §2. [38] J. Shen, R. Ren, A. Ruiz, and F. Moreno-Noguer (2024) Estimating 3d uncertainty field: quantifying uncertainty for neural radiance fields. In 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 2375–2381. Cited by: §1. [39] J. Shen, A. Ruiz, A. Agudo, and F. Moreno-Noguer (2021) Stochastic neural radiance fields: quantifying uncertainty in implicit 3d representations. In 2021 International Conference on 3D Vision (3DV), p. 972–981. Cited by: §1, §2. [40] C. Sun, M. Sun, and H. Chen (2022) Direct voxel grid optimization: super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 5459–5469. Cited by: §1. [41] N. Sünderhauf, J. Abou-Chakra, and D. Miller (2023) Density-aware nerf ensembles: quantifying predictive uncertainty in neural radiance fields. In 2023 IEEE International Conference on Robotics and Automation (ICRA), p. 9370–9376. Cited by: §1, §2, Table 1, Table 1, Table 1, §5.1. [42] M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa (2023) Nerfstudio: a modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH ’23. Cited by: §5.1. [43] D. Verbin, P. Hedman, B. Mildenhall, T. Zickler, J. T. Barron, and P. P. Srinivasan (2022) Ref-nerf: structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 5481–5490. Cited by: §2. [44] G. Wang, L. Pan, S. Peng, S. Liu, C. Xu, Y. Miao, W. Zhan, M. Tomizuka, M. Pollefeys, and H. Wang (2024) NeRF in robotics: a survey. arXiv preprint arXiv:2405.01333. Cited by: §1. [45] X. Wang, S. Hu, H. Fan, H. Zhu, and X. Li (2024) Neural radiance fields in medical imaging: challenges and next steps. arXiv preprint arXiv:2402.17797. Cited by: §1. [46] Q. Xu, Z. Xu, J. Philip, S. Bi, Z. Shu, K. Sunkavalli, and U. Neumann (2022) Point-nerf: point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 5438–5448. Cited by: §2. [47] X. Xu, S. Han, Y. Zhou, W. Yin, F. Zhang, and L. Chen (2025) ENeRF: evidential neural radiance fields with uncertainty-driven adaptive resampling for sparse-view synthesis. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), p. 468–481. Cited by: §2. [48] K. Yücer, A. Sorkine-Hornung, O. Wang, and O. Sorkine-Hornung (2016) Efficient 3d object segmentation from densely sampled light fields with applications to 3d reconstruction. ACM Transactions on Graphics (TOG) 35 (3), p. 1–15. Cited by: Table 1, §5.1. Supplementary Material 8 Derivations: Point-to-Pixel Propagation of Radiance and Uncertainty This section derives the propagation formulae of radiance and uncertainties specified in Equations 12 and 13. The mean color aggregation formula follows directly from the linearity of expectation, which does not require any independence assumption. c¯=[c]=[∑i=1Nwici]=∑i=1Nwi[ci]=∑i=1Nwic¯i. c=E[c]=E [ _i=1^Nw_ic_i ]= _i=1^Nw_iE[c_i]= _i=1^Nw_i c_i. (26) Deriving the three uncertainty propagation formulae each requires its own independence assumption. However, only two, and any two, of these three assumptions are necessary to derive all the three formulae. Assumption 1. The point colors are independent, i.e., ci⟂cj,∀i≠jc_i \!\!\! c_j,\ ∀ i≠ j. Assumption 2. The point colors are conditionally independent given the conditional means and variances, i.e., ci⟂cj∣,∀i≠jc_i \!\!\! c_j θ,\ ∀ i≠ j. Assumption 3. The point colors’ conditional means are independent, i.e., μi⟂μj,∀i≠j _i \!\!\! _j,\ ∀ i≠ j. With these assumptions, it can be shown that U U =Var[c]=Var[∑i=1Nwici]=A.1∑i=1Nwi2Var[ci]=∑i=1Nwi2Ui, =Var[c]=Var [ _i=1^Nw_ic_i ] A.1= _i=1^Nw_i^2Var[c_i]= _i=1^Nw_i^2U_i, (27) Ualea U^alea =[Var[c|]]=[Var[∑i=1Nwici|]]=A.2[∑i=1Nwi2Var[ci|]] =E[Var[c\,|\, θ]]=E [Var [ _i=1^Nw_ic_i\ |\ θ ] ] A.2=E [ _i=1^Nw_i^2Var[c_i\,|\, θ] ] =∑i=1Nwi2[Var[ci|]]=∑i=1Nwi2[Var[ci|μi,σi2]]=∑i=1Nwi2Uialea, = _i=1^Nw_i^2E [Var[c_i\,|\, θ] ]= _i=1^Nw_i^2E [Var[c_i\,|\, _i, _i^2] ]= _i=1^Nw_i^2U_i^alea, (28) Uepis U^epis =Var[[c|]]=Var[[∑i=1Nwici|]]=Var[∑i=1Nwi[ci|]] =Var[E[c\,|\, θ]]=Var [E [ _i=1^Nw_ic_i\ |\ θ ] ]=Var [ _i=1^Nw_iE[c_i\,|\, θ] ] =A.3∑i=1Nwi2Var[[ci|]]=∑i=1Nwi2Var[[ci|μi,σi2]]=∑i=1Nwi2Uiepis, A.3= _i=1^Nw_i^2Var [E[c_i\,|\, θ] ]= _i=1^Nw_i^2Var[E[c_i\,|\, _i, _i^2]]= _i=1^Nw_i^2U_i^epis, (29) where =A.n A.n= denotes the step where Assumption n is used. These three equations are connected by the law of total variance, as Var[c]⏟U Var[c]_U =[Var[c|]]⏟Ualea+Var[[c|]]⏟Uepis, = E[Var[c\,|\, θ]]_U^alea+ Var[E[c\,|\, θ]]_U^epis, (30) Var[ci]⏟Ui Var[c_i]_U_i =[Var[ci|μi,σi2]]⏟Uialea+Var[[ci|μi,σi2]]⏟Uiepis. = E[Var[c_i\,|\, _i, _i^2]]_U_i^alea+ Var[E[c_i\,|\, _i, _i^2]]_U_i^epis. (31) Therefore, any two of Equations 27, 8 and 8 imply the third, and thus only two of Assumptions 1, 2 and 3 are necessary to derive all the three equations. 9 Derivations: Pixel Radiance Marginal Distribution and Loss Function This section derives the marginal distribution of pixel radiance and the negative log-likelihood loss in Equations 21 and 23. Evidential NeRF defines each pixel radiance c by a hierarchical probabilistic model c|μ,σ2∼(μ,σ2),μ|σ2∼(γ,σ2/ν),σ2∼Γ−1(α,β),c\ |\ μ,σ^2 (μ,σ^2), μ\ |\ σ^2 (γ,σ^2/ν), σ^2 ^-1(α,β), (32) where (⋅,⋅)N(·,·) and Γ−1(⋅,⋅) ^-1(·,·) respectively denote normal distribution and inverse-gamma distribution and γ∈ℝγ , ν>0ν>0, α>1α>1, and β>0β>0 are the evidential NIG parameters. Based on these prerequisites, we derive the marginal distribution and negative log-likelihood of the pixel color c. The probability density functions of the distributions in Equation 32 are respectively given by p(c|μ,σ2) p(c\ |\ μ,σ^2) =12πσ2exp(−(c−μ)22σ2), = 1 2πσ^2 (- (c-μ)^22σ^2 ), (33) p(μ|σ2) p(μ\ |\ σ^2) =ν2πσ2exp(−(μ−γ)2ν2σ2), = ν 2πσ^2 (- (μ-γ)^2ν2σ^2 ), (34) p(σ2) p(σ^2) =βαΓ(α)(σ2)−α−1exp(−βσ2). = β^α (α)(σ^2)^-α-1 (- βσ^2 ). (35) It can therefore be shown that p(c|σ2) \,\,\,\,p(c\ |\ σ^2) (36) =∫−∞p(c|μ,σ2)p(μ|σ2)μ = _-∞^∞p(c\ |\ μ,σ^2)p(μ\ |\ σ^2)dμ (37) =∫−∞12πσ2exp(−(c−μ)22σ2)ν2πσ2exp(−(μ−γ)2ν2σ2)μ = _-∞^∞ 1 2πσ^2 (- (c-μ)^22σ^2 ) ν 2πσ^2 (- (μ-γ)^2ν2σ^2 )dμ (38) =∫−∞ν2πσ2exp(−(c−μ)2+(μ−γ)2ν2σ2)μ = _-∞^∞ ν2πσ^2 (- (c-μ)^2+(μ-γ)^2ν2σ^2 )dμ (39) =∫−∞ν2πσ2exp(−(ν+1)μ2−2(c+γν)μ+(c2+γ2ν)2σ2)μ = _-∞^∞ ν2πσ^2 (- (ν+1)μ^2-2(c+γν)μ+(c^2+γ^2ν)2σ^2 )dμ (40) =ν2πσ2exp(−c2+γ2ν2σ2)∫−∞exp(−(ν+1)μ2−2(c+γν)μ2σ2)μ = ν2πσ^2 (- c^2+γ^2ν2σ^2 ) _-∞^∞ (- (ν+1)μ^2-2(c+γν)μ2σ^2 )dμ (41) =ν2πσ2exp(−c2+γ2ν2σ2)∫−∞exp(−(μ−c+γν+1)2−(c+γν+1)22σ2ν+1)μ = ν2πσ^2 (- c^2+γ^2ν2σ^2 ) _-∞^∞ (- (μ- c+γν+1)^2-( c+γν+1)^2 2σ^2ν+1 )dμ (42) =ν2πσ2exp(−c2+γ2ν2σ2)exp((c+γν)22σ2(ν+1))∫−∞exp(−(μ−c+γν+1)22σ2ν+1)μ = ν2πσ^2 (- c^2+γ^2ν2σ^2 ) ( (c+γν)^22σ^2(ν+1) ) _-∞^∞ (- (μ- c+γν+1)^2 2σ^2ν+1 )dμ (43) =ν2πσ2exp(−c2+γ2ν2σ2+(c+γν)22σ2(ν+1))∫−∞2πσ2ν+1(μ;c+γν+1,σ2ν+1)μ = ν2πσ^2 (- c^2+γ^2ν2σ^2+ (c+γν)^22σ^2(ν+1) ) _-∞^∞ 2π σ^2ν+1N (μ; c+γν+1, σ^2ν+1 )dμ (44) =ν2πσ22πσ2ν+1exp(−(c2+γ2ν)(ν+1)−(c+γν)22σ2(ν+1))∫−∞(μ;c+γν+1,σ2ν+1)μ = ν2πσ^2 2π σ^2ν+1 (- (c^2+γ^2ν)(ν+1)-(c+γν)^22σ^2(ν+1) ) _-∞^∞N (μ; c+γν+1, σ^2ν+1 )dμ (45) =ν2πσ22πσ2ν+1exp(−(c2+γ2ν)(ν+1)−(c+γν)22σ2(ν+1)) = ν2πσ^2 2π σ^2ν+1 (- (c^2+γ^2ν)(ν+1)-(c+γν)^22σ^2(ν+1) ) (46) =ν2πσ2(ν+1)exp(−(c−γ)2ν2σ2(ν+1)), = ν2πσ^2(ν+1) (- (c-γ)^2ν2σ^2(ν+1) ), (47) i.e., c|σ2∼(γ,σ2(ν+1)ν),c\ |\ σ^2 (γ, σ^2(ν+1)ν ), (48) and thus, p(c) \,\,\,\,p(c) (49) =∫0∞p(c|σ2)p(σ2)σ2 = _0^∞p(c\ |\ σ^2)p(σ^2)dσ^2 (50) =∫0∞ν2πσ2(ν+1)exp(−(c−γ)2ν2σ2(ν+1))βαΓ(α)(σ2)−α−1exp(−βσ2)σ2 = _0^∞ ν2πσ^2(ν+1) (- (c-γ)^2ν2σ^2(ν+1) ) β^α (α)(σ^2)^-α-1 (- βσ^2 )dσ^2 (51) =ν2π(ν+1)βαΓ(α)∫0∞exp(−(c−γ)2ν2σ2(ν+1)−βσ2)(σ2)−α−32σ2 = ν2π(ν+1) β^α (α) _0^∞ (- (c-γ)^2ν2σ^2(ν+1)- βσ^2 )(σ^2)^-α- 32dσ^2 (52) =ν2π(ν+1)βαΓ(α)∫0∞exp(−(c−γ)2ν2(ν+1)+βσ2)(σ2)−α−32σ2 = ν2π(ν+1) β^α (α) _0^∞ (- (c-γ)^2ν2(ν+1)+βσ^2 )(σ^2)^-α- 32dσ^2 (53) =ν2π(ν+1)βαΓ(α)∫0∞Γ(α+12)((c−γ)2ν2(ν+1)+β)α+12Γ−1(σ2;α+12,(c−γ)2ν2(ν+1)+β)σ2 = ν2π(ν+1) β^α (α) _0^∞ (α+ 12 ) ( (c-γ)^2ν2(ν+1)+β )^α+ 12 ^-1 (σ^2;α+ 12, (c-γ)^2ν2(ν+1)+β )dσ^2 (54) =ν2π(ν+1)βαΓ(α)Γ(α+12)((c−γ)2ν2(ν+1)+β)α+12∫0∞Γ−1(σ2;α+12,(c−γ)2ν2(ν+1)+β)σ2 = ν2π(ν+1) β^α (α) (α+ 12 ) ( (c-γ)^2ν2(ν+1)+β )^α+ 12 _0^∞ ^-1 (σ^2;α+ 12, (c-γ)^2ν2(ν+1)+β )dσ^2 (55) =ν2π(ν+1)βαΓ(α)Γ(α+12)((c−γ)2ν2(ν+1)+β)α+12 = ν2π(ν+1) β^α (α) (α+ 12 ) ( (c-γ)^2ν2(ν+1)+β )^α+ 12 (56) =Γ(α+12)Γ(α)2πβ(ν+1)ν((c−γ)22β(ν+1)ν+1)−(α+12) = (α+ 12 ) (α) 2π β(ν+1)ν ( (c-γ)^22 β(ν+1)ν+1 )^- (α+ 12 ) (57) =Γ(νt+12)Γ(νt2)πνtσt2((c−μt)2νtσt2+1)−νt+12, = ( _t+12) ( _t2) π _t _t^2 ( (c- _t)^2 _t _t^2+1 )^- _t+12, (58) i.e., c∼t(μt=γ,σt2=β(ν+1)αν,νt=2α),c t ( _t=γ, _t^2= β(ν+1)αν, _t=2α ), (59) where t(μt,σt2,νt)t( _t, _t^2, _t) denotes a Student’s t distribution with location μt _t, scale σt _t, and degrees of freedom νt _t, and (x;⋅,⋅)N(x;·,·) and Γ−1(x;⋅,⋅) ^-1(x;·,·) represent their probability distribution densities at x. This yields Equation 21. The negative log-likelihood of c can thereby be derived as −logp(c) - p(c) (60) =−log(Γ(α+12)Γ(α)2πβ(ν+1)ν((c−γ)22β(ν+1)ν+1)−(α+12)) =- ( (α+ 12 ) (α) 2π β(ν+1)ν ( (c-γ)^22 β(ν+1)ν+1 )^- (α+ 12 ) ) (61) =12log(2πβ(ν+1)ν)+logΓ(α)Γ(α+12)+(α+12)log((c−γ)22β(ν+1)ν+1) = 12 (2π β(ν+1)ν )+ (α) (α+ 12 )+ (α+ 12 ) ( (c-γ)^22 β(ν+1)ν+1 ) (62) =12logπν+12log(2β(ν+1))+logΓ(α)Γ(α+12)+(α+12)log((c−γ)2ν2β(ν+1)+1) = 12 πν+ 12 (2β(ν+1) )+ (α) (α+ 12 )+ (α+ 12 ) ( (c-γ)^2ν2β(ν+1)+1 ) (63) =12logπν−αlog(2β(ν+1))+logΓ(α)Γ(α+12)+(α+12)log((c−γ)2ν+2β(ν+1)) = 12 πν-α (2β(ν+1) )+ (α) (α+ 12 )+ (α+ 12 ) ((c-γ)^2ν+2β(ν+1) ) (64) =12logπν−αlogΩ+logΓ(α)Γ(α+12)+(α+12)log((c−γ)2ν+Ω), = 12 πν-α + (α) (α+ 12 )+ (α+ 12 ) ((c-γ)^2ν+ ), (65) where Ω=2β(1+ν) =2β(1+ν). This yields Equation 23. 10 Further Discussions More quantitative results. In the main paper, quantitative results averaged across all scenes within each dataset are reported. Tables 4, 5, 6 and 7 provide per-scene statistics including the mean and standard deviation of metrics over three independent runs. More qualitative results. We present the qualitative comparison of the uncertainty methods on LF, LLFF, and RobustNeRF datasets in Figures 11, 12 and 13, respectively. Additionally, we show the aleatoric and epistemic uncertainty maps of more scenes in the wild from Phototoursim [16] in Figure 14. Hyperparameter selection. The regularization coefficient in the loss function is selected based on the quantitative metrics. Figure 9 illustrates how different coefficients affect image reconstruction and uncertainty estimation. In general, both excessively small and large coefficients lead to suboptimal performance, and the best hyperparameter value is inherently scene-dependent. The specific regularization coefficients utilized to produce the reported results are detailed in Table 3. Figure 9: Sensitivity study of regularization coefficient’s effect on all the quantitative metrics of Leaves scene. Scene Africa Basket Statue Torch Fern Flower Fortress Horns Leaves Orchids Room T-Rex Android Crab Statue Yoda λreg _reg 5e-3 1e-1 1e-3 1e0 1e0 5e-2 1e0 5e-1 5e0 1e0 5e-1 5e0 5e-3 5e-4 1e-4 1e-5 Table 3: The regularization coefficients used in each scene of LF, LLFF, and RobustNeRF datasets. Mutual causes of aleatoric and epistemic uncertainties. Various complex elements in a 3D scene can lead to elevated levels of either aleatoric or epistemic uncertainty. However, attributing each specific factor exclusively to one type of uncertainty is often inappropriate, as many factors affect both uncertainties in different ways and to varying degrees. For example, transient objects increase AU due to color variations introduced by motion, while simultaneously raising EU through partial occlusions. Similarly, edges or high-frequency non-smooth regions tend to exhibit higher AU since their radiance is highly sensitive to input rays, as small inaccuracies in sensing, digitization, or poses can yield large radiance variations, resulting in nearly irreducible data uncertainty; meanwhile, the irregular geometry of such regions obstructs ray coverage and limits supervision signals from those surfaces, thereby increasing EU as well. Uncertainties of transients. Transient objects can lead to both higher aleatoric and epistemic uncertainties. In practice, the uncertainties of the transient regions depend on the densities assigned to them by the model. Figure 10 shows ten images from two RobustNeRF scenes where the model is trained on images with cluttered objects. It can be observed that when the model cannot disambiguate the transients and the floaters appear in the test renderings, both AU and EU tend to be higher on them, meaning that the model simultaneously receives inconsistent radiance signals (high AU) and lacks sufficient knowledge to determine the presence or geometry of the transients (high EU). When the model resolves the transient objects (by minimizing their densities and removing them from volumetric rendering), the floaters disappear in the test image reconstructions and only AU remains high, indicating that the model no longer lacks the knowledge to determine the presence of the transients but still records the color inconsistency from the training signals as high AU. Figure 10: Aleatoric and epistemic uncertainties of scenes with transient objects. The red bounding boxes delineate the erroneous artifacts in test renderings caused by transients in the training views. If the model fails to suppress the floaters, both AU and EU are elevated on the transients; If the model resolves the transience, only AU is higher on the regions where the training transients were once present. Scene Method PSNR↑ SSIM↑ LPIPS↓ NLL↓ AUSE RMSE↓ AUSE MAE↓ Baseline 26.5406 ± 0.1903 0.9019 ± 0.0009 0.0565 ± 0.0010 Dropout 26.6798 ± 0.0205 0.8986 ± 0.0007 0.0590 ± 0.0014 4.0201 ± 0.2676 0.0148 ± 0.0003 0.0064 ± 0.0001 Normal 27.4766 ± 1.5843 0.9028 ± 0.0192 0.0608 ± 0.0210 -1.4911 ± 0.6587 0.0078 ± 0.0023 0.0033 ± 0.0004 Africa MoL 28.0663 ± 0.6656 0.9023 ± 0.0082 0.0614 ± 0.0056 -2.3375 ± 0.0088 0.0069 ± 0.0001 0.0035 ± 0.0001 Ensembles 27.0083 ± 0.0652 0.9171 ± 0.0003 0.0491 ± 0.0006 0.0871 ± 0.1899 0.0087 ± 0.0005 0.0037 ± 0.0001 DANE 27.0083 ± 0.0652 0.9171 ± 0.0003 0.0491 ± 0.0006 -0.9581 ± 0.1721 0.0139 ± 0.0010 0.0059 ± 0.0003 Evidential 29.8826 ± 0.0617 0.9331 ± 0.0002 0.0340 ± 0.0015 -2.3643 ± 0.0030 0.0054 ± 0.0000 0.0028 ± 0.0000 Baseline 28.0171 ± 0.0703 0.9114 ± 0.0019 0.0474 ± 0.0008 Dropout 27.3038 ± 0.0176 0.8897 ± 0.0018 0.0610 ± 0.0016 6.2637 ± 0.3747 0.0155 ± 0.0005 0.0054 ± 0.0002 Normal 27.9387 ± 0.6029 0.9087 ± 0.0097 0.0517 ± 0.0044 5.5852 ± 2.7302 0.0137 ± 0.0033 0.0043 ± 0.0013 Basket MoL 27.5847 ± 0.2379 0.9044 ± 0.0007 0.0692 ± 0.0015 -2.2614 ± 0.0438 0.0125 ± 0.0016 0.0036 ± 0.0002 Ensembles 28.9951 ± 0.0715 0.9258 ± 0.0011 0.0427 ± 0.0004 -0.8529 ± 0.3663 0.0058 ± 0.0002 0.0023 ± 0.0001 DANE 28.9951 ± 0.0715 0.9258 ± 0.0011 0.0427 ± 0.0004 -1.0245 ± 0.3048 0.0087 ± 0.0010 0.0034 ± 0.0004 Evidential 29.1442 ± 0.1841 0.9263 ± 0.0008 0.0369 ± 0.0002 -2.1757 ± 0.0901 0.0098 ± 0.0008 0.0033 ± 0.0001 Baseline 32.8018 ± 0.1611 0.9645 ± 0.0007 0.0221 ± 0.0007 Dropout 31.8005 ± 0.0787 0.9525 ± 0.0009 0.0377 ± 0.0007 -1.1505 ± 0.0358 0.0057 ± 0.0002 0.0031 ± 0.0001 Normal 28.8110 ± 1.1896 0.9457 ± 0.0036 0.0424 ± 0.0033 -1.9682 ± 0.4526 0.0055 ± 0.0010 0.0017 ± 0.0001 Statue MoL 30.6262 ± 0.5385 0.9469 ± 0.0083 0.0545 ± 0.0174 -2.9409 ± 0.0318 0.0030 ± 0.0002 0.0016 ± 0.0001 Ensembles 33.7663 ± 0.0662 0.9718 ± 0.0001 0.0198 ± 0.0000 -1.9914 ± 0.0994 0.0029 ± 0.0001 0.0017 ± 0.0000 DANE 33.7663 ± 0.0662 0.9718 ± 0.0001 0.0198 ± 0.0000 -2.1733 ± 0.0538 0.0034 ± 0.0002 0.0020 ± 0.0001 Evidential 32.8506 ± 0.2748 0.9661 ± 0.0003 0.0221 ± 0.0003 -2.9793 ± 0.0192 0.0031 ± 0.0001 0.0014 ± 0.0000 Baseline 26.8557 ± 0.0802 0.8911 ± 0.0010 0.0599 ± 0.0014 Dropout 26.6709 ± 0.0290 0.8835 ± 0.0005 0.0698 ± 0.0007 5.5346 ± 0.4652 0.0139 ± 0.0005 0.0046 ± 0.0001 Normal 27.7991 ± 0.1333 0.9087 ± 0.0019 0.0574 ± 0.0015 -0.3559 ± 1.1275 0.0090 ± 0.0022 0.0023 ± 0.0002 Torch MoL 26.6030 ± 0.1090 0.8845 ± 0.0028 0.0838 ± 0.0042 -2.6173 ± 0.0145 0.0112 ± 0.0005 0.0028 ± 0.0000 Ensembles 27.7419 ± 0.0417 0.9086 ± 0.0004 0.0528 ± 0.0002 4.0553 ± 1.4211 0.0105 ± 0.0006 0.0026 ± 0.0001 DANE 27.7419 ± 0.0417 0.9086 ± 0.0004 0.0528 ± 0.0002 2.4292 ± 0.9753 0.0144 ± 0.0013 0.0044 ± 0.0005 Evidential 27.9943 ± 0.1674 0.9126 ± 0.0017 0.0507 ± 0.0041 -2.2771 ± 0.1350 0.0096 ± 0.0022 0.0025 ± 0.0003 Table 4: Mean and standard deviation of quantitative metrics over three runs on LF. Scene Method PSNR↑ SSIM↑ LPIPS↓ NLL↓ AUSE RMSE↓ AUSE MAE↓ Baseline 20.1069 ± 0.1190 0.5685 ± 0.0005 0.3617 ± 0.0050 Dropout 19.1317 ± 0.1808 0.5153 ± 0.0059 0.4432 ± 0.0060 46.5072 ± 8.3738 0.0517 ± 0.0016 0.0294 ± 0.0020 Normal 20.1763 ± 0.3308 0.5905 ± 0.0115 0.3273 ± 0.0133 47.5151 ± 1.6352 0.0349 ± 0.0014 0.0193 ± 0.0005 Fern MoL 18.2953 ± 0.6532 0.5222 ± 0.0298 0.4279 ± 0.0406 1.3336 ± 0.2902 0.0605 ± 0.0157 0.0257 ± 0.0078 Ensembles 20.7328 ± 0.0591 0.6134 ± 0.0016 0.3462 ± 0.0019 3.2024 ± 0.4579 0.0297 ± 0.0004 0.0148 ± 0.0002 DANE 20.7328 ± 0.0591 0.6134 ± 0.0016 0.3462 ± 0.0019 2.6334 ± 0.4099 0.0332 ± 0.0002 0.0172 ± 0.0001 Evidential 20.8095 ± 0.1119 0.6216 ± 0.0074 0.3005 ± 0.0077 -0.4856 ± 0.1133 0.0301 ± 0.0009 0.0160 ± 0.0007 Baseline 18.3456 ± 0.0046 0.4792 ± 0.0068 0.4230 ± 0.0116 Dropout 18.6701 ± 0.1790 0.4990 ± 0.0089 0.3780 ± 0.0057 54.1338 ± 2.6904 0.0554 ± 0.0012 0.0341 ± 0.0012 Normal 12.6188 ± 4.7635 0.2205 ± 0.1075 0.7529 ± 0.2290 9.0113 ± 6.3991 0.0732 ± 0.0073 0.0590 ± 0.0115 Flower MoL 18.6338 ± 0.2226 0.5140 ± 0.0055 0.3665 ± 0.0141 1.4129 ± 0.1601 0.0589 ± 0.0023 0.0296 ± 0.0012 Ensembles 18.7711 ± 0.0344 0.5204 ± 0.0036 0.3993 ± 0.0050 7.7562 ± 0.2043 0.0370 ± 0.0004 0.0173 ± 0.0002 DANE 18.7711 ± 0.0344 0.5204 ± 0.0036 0.3993 ± 0.0050 7.7131 ± 0.2053 0.0371 ± 0.0004 0.0174 ± 0.0002 Evidential 19.2606 ± 0.1150 0.5254 ± 0.0126 0.3957 ± 0.0208 2.4728 ± 0.6069 0.0409 ± 0.0022 0.0238 ± 0.0015 Baseline 18.3478 ± 0.1598 0.3906 ± 0.0033 0.5240 ± 0.0069 Dropout 18.3708 ± 0.5357 0.4063 ± 0.0039 0.5204 ± 0.0077 52.6142 ± 9.0302 0.0518 ± 0.0069 0.0278 ± 0.0022 Normal 18.3146 ± 0.0871 0.4100 ± 0.0106 0.5060 ± 0.0427 62.8832 ± 20.7391 0.0515 ± 0.0036 0.0248 ± 0.0027 Fortress MoL 17.4740 ± 0.1443 0.3779 ± 0.0064 0.6090 ± 0.0253 1.8754 ± 0.5340 0.0645 ± 0.0009 0.0280 ± 0.0003 Ensembles 18.8273 ± 0.0496 0.4438 ± 0.0015 0.4517 ± 0.0023 7.5589 ± 0.8103 0.0369 ± 0.0014 0.0188 ± 0.0003 DANE 18.8273 ± 0.0496 0.4438 ± 0.0015 0.4517 ± 0.0023 7.1596 ± 0.8387 0.0377 ± 0.0017 0.0195 ± 0.0003 Evidential 18.7310 ± 0.0406 0.4264 ± 0.0065 0.4878 ± 0.0040 0.2943 ± 0.1765 0.0404 ± 0.0011 0.0208 ± 0.0005 Baseline 15.7052 ± 0.1371 0.4527 ± 0.0105 0.4451 ± 0.0111 Dropout 15.4649 ± 0.2658 0.4296 ± 0.0113 0.4729 ± 0.0074 159.8088 ± 22.3326 0.0919 ± 0.0057 0.0482 ± 0.0025 Normal 13.4923 ± 0.2916 0.2336 ± 0.0089 0.6160 ± 0.0096 22.6808 ± 7.0886 0.1130 ± 0.0047 0.0807 ± 0.0083 Horns MoL 14.5633 ± 0.1664 0.3638 ± 0.0457 0.5691 ± 0.0891 2.1617 ± 0.2982 0.1101 ± 0.0039 0.0505 ± 0.0008 Ensembles 15.9843 ± 0.0172 0.5025 ± 0.0033 0.4240 ± 0.0045 22.0785 ± 1.0690 0.0759 ± 0.0013 0.0328 ± 0.0005 DANE 15.9843 ± 0.0172 0.5025 ± 0.0033 0.4240 ± 0.0045 15.3955 ± 1.4795 0.0750 ± 0.0014 0.0330 ± 0.0006 Evidential 15.7596 ± 0.1745 0.5034 ± 0.0074 0.3878 ± 0.0075 1.8495 ± 0.1097 0.0941 ± 0.0055 0.0403 ± 0.0023 Table 5: Mean and standard deviation of quantitative metrics over three runs on LLFF. Scene Method PSNR↑ SSIM↑ LPIPS↓ NLL↓ AUSE RMSE↓ AUSE MAE↓ Baseline 13.6976 ± 0.1464 0.2557 ± 0.0155 0.4167 ± 0.0066 Dropout 13.9549 ± 0.2228 0.2423 ± 0.0248 0.4235 ± 0.0211 141.2669 ± 12.2890 0.1072 ± 0.0025 0.0618 ± 0.0024 Normal 13.3743 ± 0.0707 0.2525 ± 0.0111 0.4239 ± 0.0075 104.3151 ± 41.1687 0.1064 ± 0.0013 0.0553 ± 0.0009 Leaves MoL 12.8939 ± 0.1964 0.2408 ± 0.0179 0.4276 ± 0.0162 4.1014 ± 0.2543 0.1266 ± 0.0053 0.0674 ± 0.0044 Ensembles 14.2397 ± 0.0569 0.2960 ± 0.0069 0.4319 ± 0.0047 12.3460 ± 0.5912 0.0915 ± 0.0010 0.0471 ± 0.0007 DANE 14.2397 ± 0.0569 0.2960 ± 0.0069 0.4319 ± 0.0047 11.6017 ± 0.1579 0.0928 ± 0.0009 0.0482 ± 0.0006 Evidential 13.9301 ± 0.0437 0.3020 ± 0.0093 0.3847 ± 0.0035 0.0349 ± 0.0607 0.0809 ± 0.0009 0.0446 ± 0.0003 Baseline 14.5292 ± 0.0830 0.3036 ± 0.0057 0.3916 ± 0.0103 Dropout 13.3110 ± 0.4149 0.2151 ± 0.0202 0.4709 ± 0.0100 123.9834 ± 10.5705 0.1195 ± 0.0056 0.0762 ± 0.0037 Normal 14.9466 ± 0.1068 0.3189 ± 0.0126 0.3911 ± 0.0160 25.0742 ± 5.0559 0.0754 ± 0.0028 0.0424 ± 0.0013 Orchids MoL 13.8818 ± 0.0307 0.3242 ± 0.0031 0.3706 ± 0.0047 2.4593 ± 0.0292 0.1058 ± 0.0045 0.0506 ± 0.0021 Ensembles 14.8488 ± 0.1379 0.3250 ± 0.0081 0.3953 ± 0.0065 9.7556 ± 0.8979 0.0753 ± 0.0037 0.0409 ± 0.0017 DANE 14.8488 ± 0.1379 0.3250 ± 0.0081 0.3953 ± 0.0065 9.4356 ± 0.7865 0.0746 ± 0.0034 0.0407 ± 0.0016 Evidential 14.7012 ± 0.1548 0.3253 ± 0.0090 0.3601 ± 0.0103 0.5681 ± 0.2424 0.0843 ± 0.0041 0.0459 ± 0.0017 Baseline 19.7859 ± 0.0325 0.7196 ± 0.0037 0.3864 ± 0.0088 Dropout 19.1522 ± 0.0418 0.6704 ± 0.0097 0.4536 ± 0.0085 79.6637 ± 8.4028 0.0637 ± 0.0016 0.0342 ± 0.0006 Normal 19.8485 ± 0.2959 0.6766 ± 0.0384 0.4178 ± 0.0630 111.9164 ± 13.8320 0.0465 ± 0.0013 0.0232 ± 0.0020 Room MoL 17.6348 ± 0.7025 0.5337 ± 0.0754 0.6094 ± 0.0874 3.1864 ± 0.4821 0.0651 ± 0.0042 0.0295 ± 0.0026 Ensembles 19.9259 ± 0.0347 0.7492 ± 0.0003 0.3656 ± 0.0006 15.8667 ± 0.7134 0.0347 ± 0.0006 0.0161 ± 0.0004 DANE 19.9259 ± 0.0347 0.7492 ± 0.0003 0.3656 ± 0.0006 13.4646 ± 0.4077 0.0368 ± 0.0004 0.0174 ± 0.0005 Evidential 19.9548 ± 0.0570 0.7172 ± 0.0181 0.3660 ± 0.0299 1.2947 ± 1.1068 0.0465 ± 0.0027 0.0221 ± 0.0012 Baseline 19.7067 ± 0.0095 0.6063 ± 0.0041 0.3463 ± 0.0047 Dropout 19.0824 ± 0.1470 0.5750 ± 0.0059 0.4119 ± 0.0046 67.4407 ± 5.4653 0.0525 ± 0.0006 0.0290 ± 0.0006 Normal 19.6283 ± 0.1296 0.6055 ± 0.0176 0.3502 ± 0.0238 59.4677 ± 29.4200 0.0379 ± 0.0026 0.0189 ± 0.0008 T-Rex MoL 17.9974 ± 0.5803 0.5127 ± 0.0837 0.4935 ± 0.1085 1.4452 ± 0.3356 0.0594 ± 0.0049 0.0255 ± 0.0029 Ensembles 20.0150 ± 0.0311 0.6373 ± 0.0019 0.3317 ± 0.0028 10.7621 ± 0.1102 0.0293 ± 0.0004 0.0145 ± 0.0002 DANE 20.0150 ± 0.0311 0.6373 ± 0.0019 0.3317 ± 0.0028 10.4152 ± 0.0855 0.0295 ± 0.0004 0.0146 ± 0.0002 Evidential 19.8874 ± 0.0246 0.6329 ± 0.0048 0.3177 ± 0.0095 -0.6168 ± 0.0407 0.0455 ± 0.0004 0.0226 ± 0.0006 Table 6: Mean and standard deviation of quantitative metrics over three runs on LLFF. Scene Method PSNR↑ SSIM↑ LPIPS↓ NLL↓ AUSE RMSE↓ AUSE MAE↓ Baseline 22.8865 ± 0.0476 0.7691 ± 0.0008 0.1545 ± 0.0011 Dropout 22.6094 ± 0.0779 0.7552 ± 0.0001 0.1699 ± 0.0009 17.8212 ± 0.1534 0.0308 ± 0.0002 0.0188 ± 0.0001 Normal 23.9819 ± 0.0770 0.8146 ± 0.0013 0.1058 ± 0.0021 4.9525 ± 0.7433 0.0227 ± 0.0006 0.0152 ± 0.0003 Android MoL 22.1825 ± 0.0186 0.7395 ± 0.0002 0.1905 ± 0.0008 -0.7117 ± 0.0656 0.0273 ± 0.0006 0.0172 ± 0.0005 Ensembles 23.4866 ± 0.0304 0.8012 ± 0.0006 0.1381 ± 0.0006 8.1690 ± 0.8485 0.0265 ± 0.0001 0.0158 ± 0.0000 DANE 23.4866 ± 0.0304 0.8012 ± 0.0006 0.1381 ± 0.0006 7.4801 ± 0.8655 0.0274 ± 0.0001 0.0164 ± 0.0000 Evidential 23.8915 ± 0.0651 0.8116 ± 0.0034 0.1047 ± 0.0007 -1.1616 ± 0.0266 0.0231 ± 0.0003 0.0153 ± 0.0002 Baseline 28.6708 ± 0.2399 0.9034 ± 0.0038 0.1124 ± 0.0072 Dropout 27.5569 ± 0.3792 0.8850 ± 0.0049 0.1186 ± 0.0065 24.1365 ± 2.5685 0.0216 ± 0.0012 0.0074 ± 0.0007 Normal 27.7154 ± 0.2508 0.9062 ± 0.0016 0.0971 ± 0.0025 2.9771 ± 1.6485 0.0151 ± 0.0007 0.0062 ± 0.0004 Crab MoL 27.0696 ± 0.2685 0.8767 ± 0.0044 0.1374 ± 0.0061 -2.5692 ± 0.0274 0.0186 ± 0.0013 0.0065 ± 0.0005 Ensembles 30.0907 ± 0.0696 0.9206 ± 0.0007 0.1072 ± 0.0006 -0.3212 ± 0.4515 0.0048 ± 0.0005 0.0021 ± 0.0001 DANE 30.0907 ± 0.0696 0.9206 ± 0.0007 0.1072 ± 0.0006 -0.0123 ± 0.2259 0.0234 ± 0.0006 0.0101 ± 0.0000 Evidential 29.9324 ± 0.0145 0.9160 ± 0.0086 0.0795 ± 0.0154 -2.0303 ± 0.2399 0.0150 ± 0.0059 0.0084 ± 0.0058 Baseline 20.1787 ± 0.0379 0.7489 ± 0.0046 0.2356 ± 0.0072 Dropout 20.0340 ± 0.0841 0.7213 ± 0.0068 0.2772 ± 0.0159 35.4746 ± 2.1971 0.0464 ± 0.0013 0.0304 ± 0.0008 Normal 20.1453 ± 0.0154 0.7780 ± 0.0039 0.2128 ± 0.0061 17.9184 ± 2.0826 0.0413 ± 0.0008 0.0275 ± 0.0007 Statue MoL 19.3304 ± 0.1637 0.6834 ± 0.0032 0.3300 ± 0.0069 0.0773 ± 0.0274 0.0571 ± 0.0005 0.0380 ± 0.0007 Ensembles 20.5674 ± 0.0059 0.7843 ± 0.0008 0.2166 ± 0.0028 9.9047 ± 0.3519 0.0297 ± 0.0003 0.0191 ± 0.0001 DANE 20.5674 ± 0.0059 0.7843 ± 0.0008 0.2166 ± 0.0028 8.4305 ± 0.2763 0.0367 ± 0.0002 0.0235 ± 0.0000 Evidential 20.8284 ± 0.0496 0.8017 ± 0.0010 0.1780 ± 0.0019 -0.0313 ± 0.3149 0.0362 ± 0.0008 0.0243 ± 0.0004 Baseline 29.1458 ± 0.4079 0.8969 ± 0.0081 0.1284 ± 0.0118 Dropout 28.8770 ± 0.1990 0.8857 ± 0.0041 0.1322 ± 0.0020 14.0873 ± 5.1782 0.0148 ± 0.0034 0.0073 ± 0.0014 Normal 29.3523 ± 0.5839 0.9099 ± 0.0040 0.1088 ± 0.0043 16.1012 ± 8.0429 0.0209 ± 0.0034 0.0115 ± 0.0018 Yoda MoL 26.5671 ± 0.3613 0.8688 ± 0.0051 0.1617 ± 0.0070 -2.3751 ± 0.0676 0.0186 ± 0.0018 0.0088 ± 0.0010 Ensembles 30.6364 ± 0.1516 0.9188 ± 0.0025 0.1135 ± 0.0029 0.7713 ± 0.1332 0.0044 ± 0.0001 0.0020 ± 0.0001 DANE 30.6364 ± 0.1516 0.9188 ± 0.0025 0.1135 ± 0.0029 0.5383 ± 0.2742 0.0257 ± 0.0015 0.0116 ± 0.0006 Evidential 30.2646 ± 0.1857 0.9272 ± 0.0065 0.0826 ± 0.0124 -1.8577 ± 0.2073 0.0139 ± 0.0009 0.0072 ± 0.0012 Table 7: Mean and standard deviation of quantitative metrics over three runs on RobustNeRF. Figure 11: Qualitative comparison on LF. Figure 12: Qualitative comparison on LLFF. Figure 13: Qualitative comparison on RobustNeRF. Figure 14: Aleatoric and epistemic uncertainty maps for three in-the-wild scenes from Phototourism. Aleatoric uncertainty peaks in regions with high radiance variance (e.g., sky higher above, building facades, moving figures), whereas epistemic uncertainty is concentrated in areas frequently occluded (e.g., sky directly behind the buildings or objects obscured by pedestrians).