Paper deep dive
Open-World Semantic Segmentation with Sensitivity Modeling
Anastasios Romanos Varvarigos, Nikos Giakoumoglou, Tania Stathaki
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/12/2026, 1:21:23 AM
Summary
The paper proposes a three-decoder architecture for open-world semantic segmentation to address the limitations of closed-world models in detecting anomalous or unseen content. The architecture includes a semantic decoder for known classes using Gaussian prototypes, a contrastive decoder for isolating unknown regions via feature norms, and a novel sensitivity decoder that captures fine-grained texture irregularities and activation instabilities. Experiments on Cityscapes and BDD-Anomaly datasets demonstrate that this approach improves anomaly segmentation and novel-class discovery while maintaining competitive closed-set accuracy, achieving a +2.4% AUROC gain on BDD-Anomaly over the baseline.
Entities (10)
Relation Signals (8)
Proposed Method → evaluatedon → Cityscapes
confidence 98% · Experiments on Cityscapes and BDD-Anomaly show that our method improves anomaly segmentation
Proposed Method → evaluatedon → BDD-Anomaly
confidence 98% · Experiments on Cityscapes and BDD-Anomaly show that our method improves anomaly segmentation
Sensitivity Decoder → captures → fine-grained texture irregularities
confidence 95% · The third, our key contribution, is a sensitivity decoder that captures fine-grained texture irregularities and activation instabilities
Contrastive Decoder → uses → contrastive feature learning
confidence 95% · The second uses contrastive feature learning to isolate unknown regions in embedding space.
Semantic Decoder → uses → Gaussian Prototypes
confidence 95% · The first decoder performs closed-set segmentation using Gaussian prototypes for known categories.
Sensitivity Decoder → complements → Semantic Decoder
confidence 92% · The three decoders provide genuinely complementary signals
Sensitivity Decoder → complements → Contrastive Decoder
confidence 92% · The three decoders provide genuinely complementary signals
Proposed Method → outperforms → ContMAV
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Modern vision systems must operate in "open-world" settings, where models must recognize known categories and detect unseen or anomalous content. Conventional semantic segmentation models operate under a "closed-world" assumption, often producing overconfident misclassifications on novel content. We address open-world semantic segmentation, the joint task of segmenting known classes while detecting and grouping novel or anomalous content without additional supervision, by extending a dual-decoder baseline with a third, complementary decoder within a unified encoder-decoder design. The first decoder performs closed-set segmentation using Gaussian prototypes for known categories. The second uses contrastive feature learning to isolate unknown regions in embedding space. The third, our key contribution, is a sensitivity decoder that captures fine-grained texture irregularities and activation instabilities indicative of semantic uncertainty, which neither semantic prototypes nor contrastive norms can reliably detect. The three decoders provide genuinely complementary signals: class-level OOD distance in logit space, global feature energy in embedding space, and local activation instability across encoder scales. Experiments on Cityscapes and BDD-Anomaly show that our method improves anomaly segmentation and novel-class discovery while maintaining competitive closed-set accuracy, with gains of +2.4% AUROC and a 2.5 pp. reduction in FPR@95TPR on BDD-Anomaly over the baseline.
Tags
Links
- Source: https://arxiv.org/abs/2608.08308v1
- Canonical: https://arxiv.org/abs/2608.08308v1
Trouble viewing inline? Open PDF directly →
Full Text
29,049 characters extracted from source content.
Expand or collapse full text
Open-World Semantic Segmentation with Sensitivity Modeling Abstract Modern vision systems must operate in “open-world” settings, where models must recognize known categories and detect unseen or anomalous content. Conventional semantic segmentation models operate under a “closed-world” assumption, often producing overconfident misclassifications on novel content. We address open-world semantic segmentation, the joint task of segmenting known classes while detecting and grouping novel or anomalous content without additional supervision, by extending a dual-decoder baseline with a third, complementary decoder within a unified encoder-decoder design. The first decoder performs closed-set segmentation using Gaussian prototypes for known categories. The second uses contrastive feature learning to isolate unknown regions in embedding space. The third, our key contribution, is a sensitivity decoder that captures fine-grained texture irregularities and activation instabilities indicative of semantic uncertainty, which neither semantic prototypes nor contrastive norms can reliably detect. The three decoders provide genuinely complementary signals: class-level OOD distance in logit space, global feature energy in embedding space, and local activation instability across encoder scales. Experiments on Cityscapes and BDD-Anomaly show that our method improves anomaly segmentation and novel-class discovery while maintaining competitive closed-set accuracy, with gains of +2.4% AUROC and a 2.5 p. reduction in FPR@95TPR on BDD-Anomaly over the baseline. Index Terms— Semantic Segmentation, Open-World Learning, Anomaly Detection, Uncertainty Estimation 1 Introduction Semantic segmentation assigns a semantic label (e.g., road, pedestrian, car) to every pixel of an image. Despite significant progress, most segmentation models operate under a “closed-world” assumption: they are trained and evaluated on a fixed, predefined set of classes. This assumption fails in real-world deployments, where models frequently encounter out-of-distribution (OOD) content such as unexpected road debris, novel vehicle types, or context-dependent anomalies. Conventional networks tend to produce overconfident misclassifications [12], leading to unsafe behavior in safety-critical applications such as autonomous navigation and robotic scene understanding. “Open-world” semantic segmentation addresses the problem of jointly segmenting known classes and detecting novel or anomalous regions without additional supervision [2]. It combines two complementary sub-tasks: anomaly segmentation, which produces a binary map distinguishing known from unknown pixels, and novel class discovery, which performs pixel-wise classification of novel samples into coherent, previously unseen categories. Solving both simultaneously requires a model to be simultaneously discriminative (for known classes) and appropriately uncertain (for unknown content), objectives that are in tension under standard training. Existing dual-decoder approaches [19] address these objectives with two complementary branches: a semantic branch that models known class distributions in logit space, and a contrastive branch that structures embedding norms to separate known and unknown features globally. However, both operate at a relatively coarse level of abstraction: prototype distance and embedding norm are scalar, globally aggregated quantities per pixel. In practice, anomalous content often appears with subtle local cues, i.e., irregular texture patterns, atypical activation frequency content, and inconsistencies at object boundaries, that these global signals are not designed to capture. A partially occluded novel object, for instance, may produce a feature norm within the range of known classes while exhibiting locally inconsistent activation patterns that a spatially sensitive detector would flag. Similarly, road surface anomalies such as oil spills or debris may span only a few pixels, lie far from any class prototype boundary, and yet produce characteristic high-frequency activation signatures at intermediate encoder scales, cues that are washed out by the global pooling and hierarchical upsampling operations that the semantic and contrastive decoders rely on. These observations motivate a third, independent signal specifically tuned to local activation analysis. In this paper, we propose a three-decoder architecture (Section 3) that introduces a third, orthogonal objective: sensitivity modeling. The semantic decoder performs closed-set segmentation using Gaussian prototypes. The contrastive decoder separates known and unknown pixels in feature space via prototype alignment and objectosphere loss. The sensitivity decoder models local, fine-grained uncertainty arising from activation irregularities and texture anomalies that the other two branches cannot detect. All three outputs are fused into a unified segmentation map. We evaluate on BDD-Anomaly [3] and Cityscapes [5], reporting both closed-set mIoU and dedicated open-world metrics (AUROC, AUPR, FPR@95TPR), showing consistent improvements over the baseline [19] (Tabs. 1, 2 and 3). Our contributions are: (i) A sensitivity decoder for fine-grained uncertainty modeling, specifically designed to detect ambiguous boundaries and texture anomalies that semantic and contrastive branches cannot resolve, and conceptually distinct from both prototype-based and norm-based uncertainty. (i) A fully-convolutional three-decoder architecture with complementary losses, trained end-to-end on a single consumer GPU, achieving improved open-world segmentation. (i) Evaluation with closed-set and open-world metrics (AUROC, AUPR, FPR@95TPR) demonstrating consistent gains, with ablation studies validating each architectural component. 2 Related Work 2.1 Anomaly Detection and OOD Recognition Open-world segmentation builds on a long line of anomaly detection work, which addresses the realistic assumption that test-time inputs may contain pixels from classes absent at training time and asks a model to flag such pixels as unknown rather than force them into the closed label set. Early methods rely on softmax confidence thresholding [12] or energy-based scoring [16] to flag OOD inputs at the image level. Feature-space methods compare per-pixel embeddings to learned class prototypes, treating distance as an unknownness proxy [6]. These approaches share a common limitation: they operate on globally aggregated statistics per pixel, which can miss locally inconsistent but globally plausible content. 2.2 Pixel-Level Open-World Segmentation Pixel-level extensions of anomaly detection use synthetic OOD data generation [3], or reconstruction error from generative models. The closest prior work, ContMAV [19], formalizes open-world semantic segmentation with a dual-decoder design: one branch for closed-set prediction and one for contrastive anomaly detection via objectosphere loss [6]. RAML [7] and DMLNet [1] both build on a heavier DeepLabV3+ backbone trained with a multi-stage SGD pipeline, closed-set pretraining followed by pixel-level metric learning, with RAML further fine-tuning on a handful of labeled novel-class examples; their closed-set gains therefore come with a larger backbone, a longer multi-stage schedule, and additional supervision that our single-stage, fully unsupervised, end-to-end training on one consumer GPU does not require. Vision-language models have been applied to open-vocabulary anomaly segmentation [14], offering strong zero-shot generalization but relying on billion-parameter backbones incompatible with real-time deployment. 2.3 Positioning of Our Approach Our sensitivity decoder is inspired by TruFor [10], which localizes image tampering through learned local texture sensitivity maps. The underlying intuition is that anomalous content leaves high-frequency traces in intermediate activations even when global statistics appear normal. We adapt this idea to open-world segmentation, training the decoder on proxy unknown labels as a lightweight parallel head. This is conceptually distinct from prototype distance (semantic decoder), embedding norm (contrastive decoder), Bayesian uncertainty, and reconstruction error: it captures local activation instability in a single deterministic forward pass with negligible overhead. 3 Methodology 3.1 Architecture Our architecture uses a shared encoder fencf_enc and three parallel decoder heads: fsemf_sem, fconf_con, and fsenf_sen (Figure 1). Given input ∈ℝH×W×3x ^H× W× 3, the encoder produces a hierarchy of feature maps, of which the deepest is =fenc()∈ℝh×w×dz=f_enc(x) ^h× w× d, while intermediate feature maps at multiple resolutions are passed to decoders via skip connections. Fig. 1: Proposed architecture. A shared ResNet-based encoder with pyramid pooling extracts multi-scale features, which are fed to three parallel decoders: fsemf_sem for semantic segmentation with Gaussian prototype modeling, fconf_con for contrastive anomaly detection, and fsenf_sen for sensitivity-based uncertainty estimation. 3.2 Semantic Decoder fsemf_sem produces per-pixel logits sem∈ℝH×W×Ky_sem ^H× W× K for K known classes. We maintain class-wise Gaussian prototypes (k,k2)k=1K\( μ_k, σ_k^2)\_k=1^K computed from correctly classified pixels [9]: k(e) μ_k^(e) =1|Ω^k|∑p∈Ω^ksem,p,k2,(e)=1|Ω^k|∑p∈Ω^k(sem,p−k(e))2, = 1| _k| _p∈ _ky_sem,p, σ_k^2,(e)= 1| _k| _p∈ _k(y_sem,p- μ_k^(e))^2, (1) where Ω^k=p∣y^p=yp=k _k=\p y_p=y_p=k\ is the set of correctly classified pixels for class k at epoch e. Prototypes are updated at the end of each epoch using an exponential moving average with momentum 0.90.9 for training stability. The feature alignment loss penalizes deviations from class prototypes, normalized by per-class standard deviation for scale invariance: ℒfeat=1|Ω|∑p∈Ω‖sem,p−yp(e−1)‖2‖yp(e−1)‖2+ϵ.L_feat= 1| | _p∈ \|y_sem,p- μ_y_p^(e-1)\|_2\| σ_y_p^(e-1)\|_2+ε. (2) The multi-loss semantic objective combines class-weighted cross-entropy ℒCEL_CE, focal loss ℒfocalL_focal [15], soft dice loss ℒdiceL_dice, and prototype alignment ℒfeatL_feat: ℒsem=λCEℒCE+λfocalℒfocal+λdiceℒdice+λfeatℒfeat.L_sem= _CEL_CE+ _focalL_focal+ _diceL_dice+ _featL_feat. (3) This combination is motivated by the severe class imbalance inherent in urban driving datasets: cross-entropy provides stable gradient flow across all classes, focal loss downweights easy background pixels to focus learning on rare foreground and boundary pixels, and dice loss directly optimizes region overlap, complementing pixel-wise losses on small or infrequent objects. In practice, we found that the focal and dice terms together provide sufficient gradient signal for both common and rare classes, and that retaining ℒCEL_CE alongside them slightly degraded performance in our final configuration; we therefore set λCE=0 _CE=0 in the full model (cf. Section 4.4, Rows G) and H)), while listing it explicitly in ℒsemL_sem for completeness, since intermediate ablation rows do include it. 3.3 Contrastive Decoder fconf_con produces per-pixel embeddings con∈ℝH×W×dey_con ^H× W× d_e. The contrastive loss operates on ℓ2 _2-normalized embeddings ¯con=con/‖con‖2 y_con=y_con/\|y_con\|_2, aligning per-class mean normalized features with historical normalized class prototypes on the unit hypersphere [8]: ℒcont=−∑k=1Klogexp(¯con,k⊤¯k(e−1)/τ)∑i=1Kexp(¯con,k⊤¯i(e−1)/τ),L_cont=- _k=1^K ( y_con,k μ_k^(e-1)/τ) _i=1^K ( y_con,k μ_i^(e-1)/τ), (4) where ¯con,k y_con,k is the mean normalized embedding for ground-truth class k pixels and τ=0.07τ=0.07 is temperature. The objectosphere loss [6], by contrast, operates on the raw (unnormalized) embeddings cony_con, enforcing radial separation: known pixels (ΩK _K) are pushed to have embedding norm ≥ξ≥ξ, while unknown pixels (ΩU _U) are pulled toward the origin: ℒobj=1|Ω|[∑p∈ΩKmax(0,ξ−‖con,p‖2)+∑p∈ΩU‖con,p‖2].L_obj= 1| |\! [ _p∈ _K\! (0,ξ-\|y_con,p\|_2)+ _p∈ _U\|y_con,p\|_2 ]. (5) This creates a geometrically interpretable embedding space where the norm of the unnormalized con,py_con,p directly serves as an unknownness proxy, with the anomaly score at inference given by ap(c)=max(0,1−‖con,p‖2/ξ)a_p^(c)= (0,1-\|y_con,p\|_2/ξ). The contrastive objective is ℒanom=λcontℒcont+λobjℒobjL_anom= _contL_cont+ _objL_obj. 3.4 Sensitivity Decoder The semantic and contrastive decoders operate on global semantic predictions and feature norms, respectively. However, anomalous regions often manifest locally as subtle texture irregularities, unstable activations near object boundaries, or partially visible novel objects, phenomena that neither semantic prototype distance nor embedding norm can reliably detect. A partially occluded anomalous object may produce a feature norm within the range of known classes while exhibiting spatially inconsistent activation patterns that a local detector would identify. The sensitivity decoder fsenf_sen is specifically designed to capture these fine-grained, local cues. Inspired by forensic sensitivity networks [10], fsenf_sen processes encoder feature maps from three pyramid levels independently. At each scale, a shared lightweight head comprising three 1×11×1 convolutional layers with layer normalization and ReLU activations is applied. The resulting scale-wise sensitivity maps are upsampled to full resolution and fused via element-wise summation, followed by a sigmoid activation to produce the per-pixel uncertainty map sen∈[0,1]H×Wy_sen∈[0,1]^H× W. This multi-scale design preserves high-frequency spatial information that the upsampling operations in fsemf_sem and fconf_con tend to smooth out, making fsenf_sen sensitive to fine-grained local inconsistencies at each resolution level. Critically, fsenf_sen learns where the model should be uncertain rather than what class a pixel belongs to or how far a feature is from known prototypes. The three decoders therefore provide genuinely complementary signals: semantic prototype distance captures class-level OOD in logit space, contrastive norm captures global feature energy in embedding space, and sensitivity captures local activation instability across encoder scales. The decoder is trained via binary cross-entropy against proxy unknown labels up∈0,1u_p∈\0,1\, generated from void annotations in the dataset and from OOD patches synthesized via CutPaste augmentation applied to training images: ℒsens=−1|Ω|∑p∈Ω[uplogsen,p+(1−up)log(1−sen,p)].L_sens=- 1| | _p∈ \! [u_p _sen,p+(1-u_p) (1-y_sen,p) ]. (6) 3.5 Fusion and Inference The semantic confidence is derived from the Gaussian likelihood sp(sem)=maxk(sem,p∣k,diag(k2))s_p^(sem)= _kN(y_sem,p μ_k,diag( σ_k^2)), yielding semantic unknownness sunk,p(sem)=1−sp(sem)s_unk,p^(sem)=1-s_p^(sem). The contrastive anomaly score and sensitivity map are averaged: sunk,p(cont)=12(ap(c)+sen,p)s_unk,p^(cont)= 12(a_p^(c)+y_sen,p). The final unknownness score is: sunk,p=12(sunk,p(sem)+sunk,p(cont)).s_unk,p= 12(s_unk,p^(sem)+s_unk,p^(cont)). (7) Before fusion, each score is min-max normalized per image to bring all three signals to a common [0,1][0,1] range, preventing any single branch from dominating due to scale differences. Pixels with sunk,p>δs_unk,p>δ are labeled as unknown; the threshold δ=0.99δ=0.99 is selected based on [19]. Remaining pixels adopt the semantic decoder label y^p y_p. The full network is trained end-to-end with ℒtotal=ℒsem+ℒanom+ℒsensL_total=L_sem+L_anom+L_sens. 4 Experiments 4.1 Implementation Details The encoder is ResNet-34 [11] with basic residual blocks replaced by NonBottleneck-1D blocks [18] for computational efficiency, augmented with Squeeze-and-Excitation modules [13] and a Pyramid Pooling Module [20]. The semantic and contrastive decoders share a hierarchical architecture of three cascaded modules, each comprising a 3×33×3 convolution, NonBottleneck-1D blocks, 2×2× bilinear upsampling, and encoder skip connections via element-wise addition. The sensitivity decoder uses a lightweight multi-scale 1×11×1 convolutional MLP applied independently at three encoder scales, with output channels 64, 32, and 16 at the three pyramid levels, before sigmoid-gated element-wise summation. The full model has approximately 28M parameters, of which the sensitivity decoder contributes fewer than 1M (<<3.6% of the total), making it a negligible computational addition over the dual-decoder baseline. Inference speed is 24 fps at 1024×5121024×512 resolution on a single NVIDIA L40s GPU, compared to 23 fps for the dual-decoder baseline, confirming the minimal overhead introduced by fsenf_sen. 4.2 Datasets and Training Setup We evaluate on Cityscapes [5] (2,975 train / 500 val, 19 classes, closed-world evaluation) and BDD-Anomaly [3] (1,935 val images with anomalous objects under diverse weather conditions including rain, night, and fog). Images are resized to 1024×5121024×512 and normalized with ImageNet statistics. We train for 500 epochs using Adam with initial learning rate 4×10−34×10^-3, polynomial decay schedule with power 0.9, batch size 8, and weight decay 10−410^-4. Standard augmentations include random cropping, horizontal flipping, color jitter, and multi-scale resizing. Loss weights are set to λCE=0 _CE=0, λfocal=1.0 _focal=1.0, λdice=1.0 _dice=1.0, λfeat=0.5 _feat=0.5, λcont=1.0 _cont=1.0, λobj=0.5 _obj=0.5 in the final model; the objectosphere radius is ξ=1.0ξ=1.0 and the contrastive temperature is τ=0.07τ=0.07. 4.3 Main Results Results on Cityscapes (Tab. 1) and BDD-Anomaly (Tabs. 2 and 3) show consistent gains over the reproduced ContMAV baseline across all metrics.111“–” indicates metrics not reported by the original method. Entries marked (repr.) are our reproductions under a unified training setup; the gap to the original ContMAV reflects differences in schedule, augmentation, and hardware. On Cityscapes, the sensitivity decoder and multi-loss landscape provide regularization that benefits closed-set segmentation without degrading it. On BDD-Anomaly, the Recall gain confirms fewer missed detections under challenging conditions, while the AUPR improvement under class imbalance confirms that fsenf_sen provides a signal complementary to the contrastive branch. RAML [7] uses a substantially heavier backbone and multi-stage pipeline; its number is reported for reference only. Qualitative comparisons are shown in Figures 2 and 3. Fig. 2: Qualitative results on Cityscapes. Columns: input image, semantic prediction, anomaly logits, anomaly prediction, open-world prediction, semantic GT, anomaly GT. Rows: Our method (bottom) vs. ContMAV [19] (repr.) (top). Table 1: Closed-set performance on the Cityscapes validation set. mIoU (%), Recall (%), Precision (%). Method mIoU Recall Precision Closed-World Methods FCN-8s [17] 65.30 – – DeepLabv2 [4] 70.40 – – Open-World Methods DMLNet [1] 69.20 – – RAML [7] 76.70 – – ContMAV [19] 71.10 – – ContMAV (repr.) 69.28 89.6 69.7 Ours 69.88 90.2 70.4 Fig. 3: Qualitative results on BDD-Anomaly. Columns and rows are as in Figure 2. Table 2: Closed-set performance on BDD-Anomaly. mIoU (%), Recall (%), Precision (%). Method mIoU Recall Precision ContMAV [19] 62.28 82.5 63.1 ContMAV (repr.) 56.12 78.9 60.4 Ours 56.68 81.3 62.0 Table 3: Open-world anomaly detection on BDD-Anomaly. AUROC (%), AUPR (%), FPR@95TPR (%; lower is better). Method AUROC AUPR FPR@95TPR ↓ ContMAV (repr.) 63.4 41.8 67.4 Ours 65.8 44.3 64.9 4.4 Ablation Study Tab. 4 presents an incremental ablation on Cityscapes. Gaussian prototype alignment ℒfeatL_feat (Row C) is the most impactful component, providing a +4% gain: normalizing by per-class standard deviation is key, as otherwise the loss is dominated by high-variance classes and regularizes compact categories more weakly. Row D recovers the reproduced ContMAV baseline, validating our implementation. The sensitivity decoder (Row F) further improves over the dual-decoder setup (Row E), with +0.5% AUROC and 1.2 p. lower FPR@95TPR on BDD-Anomaly, confirming that fsenf_sen mainly benefits unknown-region detection rather than closed-set accuracy. Rows G and H show that replacing ℒCEL_CE with focal + dice yields a small but consistent gain: once focal loss is active, cross-entropy becomes partially redundant and its uniform weighting slightly biases the optimizer toward over-represented classes, so removing it lets focal/dice shape the loss landscape effectively. All contributions are additive: the monotonic improvement from Row A to Row H confirms that each component addresses a distinct aspect of the open-world problem. Table 4: Ablation on Cityscapes validation set. Configuration H (full model) achieves the highest performance. ℒCEL_CE ℒobjL_obj ℒfeatL_feat ℒcontL_cont ℒdiceL_dice ℒsensL_sens ℒfocalL_focal mIoU A ✓ 63.21 B ✓ ✓ 64.21 C ✓ ✓ ✓ 68.12 D ✓ ✓ ✓ ✓ 69.28 E ✓ ✓ ✓ ✓ ✓ 69.40 F ✓ ✓ ✓ ✓ ✓ ✓ 69.63 G ✓ ✓ ✓ ✓ ✓ 69.70 H ✓ ✓ ✓ ✓ ✓ ✓ 69.88 5 Conclusion We proposed a three-decoder architecture for open-world semantic segmentation, the joint task of segmenting known classes while detecting and grouping novel or anomalous content without additional supervision, extending the dual-decoder ContMAV baseline with a sensitivity decoder that provides a complementary uncertainty signal alongside prototype-based OOD distance and contrastive embedding norm. The sensitivity decoder is deterministic, lightweight (<<1M parameters), and trained on proxy labels from standard annotations. Experiments on Cityscapes and BDD-Anomaly confirm consistent gains in both closed-set mIoU and open-world metrics (AUROC, AUPR, FPR@95TPR), with ablations validating the additive contribution of each component. Future work will explore learned fusion strategies and richer OOD synthesis for the sensitivity branch. References [1] J. Cen, P. Yun, J. Cai, M. Y. Wang, and M. Liu (2021) Deep metric learning for open world semantic segmentation. External Links: Link, 2108.04562 Cited by: §2.2, Table 1. [2] R. Chan, K. Lis, S. Uhlemeyer, H. Blum, S. Honari, R. Siegwart, P. Fua, M. Salzmann, and M. Rottmann (2021) SegmentMeIfYouCan: a benchmark for anomaly segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, J. Vanschoren and S. Yeung (Eds.), Vol. 1, p. . External Links: Link Cited by: §1. [3] R. Chan, M. Rottmann, and H. Gottschalk (2021) Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Vol. , p. 5108–5117. External Links: Document Cited by: §1, §2.2, §4.2. [4] L. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille (2018) DeepLab: semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence 40 (4), p. 834–848. External Links: Document Cited by: Table 1. [5] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele (2016-06) The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §1, §4.2. [6] A. R. Dhamija, M. Günther, and T. E. Boult (2018) Reducing network agnostophobia. External Links: Link, 1811.04110 Cited by: §2.1, §2.2, §3.3. [7] H. Dong, Z. Chen, M. Yuan, Y. Xie, J. Zhao, F. Yu, B. Dong, and L. Zhang (2022) Region-aware metric learning for open world semantic segmentation via meta-channel aggregation. External Links: Link, 2205.08083 Cited by: §2.2, §4.3, Table 1. [8] N. Giakoumoglou, T. Stathaki, and A. Gkelias (2025) A review on discriminative self-supervised learning methods in computer vision. External Links: Link, 2405.04969 Cited by: §3.3. [9] N. Giakoumoglou and T. Stathaki (2025) Cluster Contrast for Unsupervised Visual Representation Learning. In Proceedings of the IEEE International Conference on Image Processing (ICIP), p. 133–138. Cited by: §3.2. [10] F. Guillaro, D. Cozzolino, A. Sud, N. Dufour, and L. Verdoliva (2023-06) TruFor: leveraging all-round clues for trustworthy image forgery detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 20606–20615. Cited by: §2.3, §3.4. [11] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 770–778. External Links: Document Cited by: §4.1. [12] D. Hendrycks and K. Gimpel (2017) A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, External Links: Link Cited by: §1, §2.1. [13] J. Hu, L. Shen, and G. Sun (2018) Squeeze-and-excitation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 7132–7141. External Links: Document Cited by: §4.1. [14] J. Jeong, Y. Zou, T. Kim, D. Zhang, A. Ravichandran, and O. Dabeer (2023-06) WinCLIP: zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 19606–19616. Cited by: §2.2. [15] T. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár (2020) Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence 42 (2), p. 318–327. External Links: Document Cited by: §3.2. [16] W. Liu, X. Wang, J. Owens, and Y. Li (2020) Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems. Cited by: §2.1. [17] J. Long, E. Shelhamer, and T. Darrell (2015) Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 3431–3440. External Links: Document Cited by: Table 1. [18] E. Romera, J. M. Álvarez, L. M. Bergasa, and R. Arroyo (2018) ERFNet: efficient residual factorized convnet for real-time semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 19 (1), p. 263–272. External Links: Document Cited by: §4.1. [19] M. Sodano, F. Magistri, L. Nunes, J. Behley, and C. Stachniss (2024) Open-World Semantic Segmentation Including Class Similarity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: §1, §1, §2.2, §3.5, Figure 2, Figure 2, Table 1, Table 2. [20] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia (2017) Pyramid scene parsing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 6230–6239. External Links: Document Cited by: §4.1.