Paper deep dive
TriCCOT: Tri-part Convolutional Conformal Transformer for Onboard Space Object Detection
Adrien Dorise, Marjorie Bellizzi, Julia Cohen, Stéphane May
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 9/14/2026, 5:33:05 AM
Summary
The paper introduces TriCCOT, a tri-part architecture for onboard space object detection that combines a convolutional Region Proposal Network (RPN), a conformal prediction stage for robust bounding box enlargement, and Aper-GATES, a hardware-friendly attention-based classifier. Aper-GATES reformulates self-attention using convolutional projections and global channel statistics to avoid quadratic complexity, enabling deployment on Xilinx Versal VCK190 FPGAs. The model demonstrates improved robustness to spatial blur and noise on DIOR and VDVRaw datasets compared to existing FPGA-compatible architectures.
Entities (9)
Relation Signals (8)
TriCCOT â evaluatedon â DIOR
confidence 99% · Experiments on the DIOR and VDVRaw datasets demonstrate competitive detection performance
TriCCOT â evaluatedon â VDVRaw
confidence 99% · Experiments on the DIOR and VDVRaw datasets demonstrate competitive detection performance
TriCCOT â containscomponent â Region Proposal Network
confidence 98% · TriCCOT combines a convolutional region proposal network, a conformal prediction stage, and Aper-GATES
TriCCOT â containscomponent â Aper-GATES
confidence 98% · TriCCOT combines a convolutional region proposal network, a conformal prediction stage, and Aper-GATES
Marjorie Bellizzi â affiliatedwith â IRT Saint-ExupĂ©ry
confidence 95% · Marjorie Bellizzi ... IRT Saint-Exupéry
Adrien Dorise â affiliatedwith â CNES
confidence 95% · Adrien Dorise ... Centre National dâEtudes Spatiales, CNES
Aper-GATES â optimizedfor â Xilinx Versal VCK190
confidence 95% · Aper-GATES... avoiding standard transformer operations that are poorly suited to CNN-oriented accelerators... full deployment on a Xilinx Versal VCK190 FPGA
TriCCOT â usesmethod â Conformal Prediction
confidence 95% · subsequently enlarged via conformal prediction
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Onboard object detection in Earth observation is constrained by limited computational resources and the absence of fully corrected imagery. While convolutional detectors are hardware-efficient, they often struggle to extract robust representations from raw and noisy data. Conversely, transformer-based models provide stronger global reasoning capabilities but remain difficult to deploy on FPGA accelerators due to quadratic attention complexity and non-compatible operations. We introduce TriCCOT, a tri-part architecture for robust and deployable onboard object detection. TriCCOT combines a convolutional region proposal network, a conformal prediction stage, and Aper-GATES, our hardware-friendly attention-based classifier. The region proposal network generates candidate bounding boxes, which are subsequently enlarged via conformal prediction, providing a distribution-free probabilistic coverage guarantee. The resulting crops are processed by Aper-GATES, which reformulates self-attention through convolutional projections, global channel statistics, and hardware-friendly gating operations, avoiding standard transformer operations that are poorly suited to CNN-oriented accelerators. Experiments on the DIOR and VDVRaw datasets demonstrate competitive detection performance and improved robustness to spatial blur and signal-dependent noise when compared to FPGA-compatible architectures. Finally, we report full deployment on a Xilinx Versal VCK190 FPGA without modifying the underlying DPU architecture, enabling unified CNN-Transformer inference for spaceborne embedded applications.
Tags
Links
- Source: https://arxiv.org/abs/2609.08659v1
- Canonical: https://arxiv.org/abs/2609.08659v1
Trouble viewing inline? Open PDF directly â
Full Text
55,030 characters extracted from source content.
Expand or collapse full text
arXiv:2609.08659v1 [cs.CV] 8 Sep 2026 DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION1 TriCCOT: Tri-part Convolutional Conformal Transformer for Onboard Space Object Detection Adrien Dorise 1,2 adrien.dorise@cnes.fr Marjorie Bellizzi 2 marjorie.bellizzi@irt-saintexupery.com Julia Cohen 1 julia.cohen@cnes.fr StĂ©phane May 1 stephane.may@cnes.fr 1 Centre National dâEtudes Spatiales, CNES Toulouse, France 2 IRT Saint-ExupĂ©ry Toulouse, France Abstract Onboard object detection in Earth observation is constrained by limited computa- tional resources and the absence of fully corrected imagery. While convolutional de- tectors are hardware-efficient, they often struggle to extract robust representations from raw and noisy data. Conversely, transformer-based models provide stronger global rea- soning capabilities but remain difficult to deploy on FPGA accelerators due to quadratic attention complexity and non-compatible operations. We introduce TriCCOT, a tri-part architecture for robust and deployable onboard object detection. TriCCOT combines a convolutional region proposal network, a con- formal prediction stage, and Aper-GATES, our hardware-friendly attention-based classi- fier. The region proposal network generates candidate bounding boxes, which are sub- sequently enlarged via conformal prediction, providing a distribution-free probabilistic coverage guarantee. The resulting crops are processed by Aper-GATES, which refor- mulates self-attention through convolutional projections, global channel statistics, and hardware-friendly gating operations, avoiding standard transformer operations that are poorly suited to CNN-oriented accelerators. Experiments on the DIOR and VDVRaw datasets demonstrate competitive detection performance and improved robustness to spatial blur and signal-dependent noise when compared to FPGA-compatible architectures. Finally, we report full deployment on a Xilinx Versal VCK190 FPGA without modifying the underlying DPU architecture, en- abling unified CNN-Transformer inference for spaceborne embedded applications. 1 Introduction Earth observation systems rely extensively on image processing techniques to transform raw, noisy acquisitions into exploitable products delivered to end users. Radiometric and geo- metric corrections are traditionally performed on the ground after data transmission from © 2026. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms. 2DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION the satellite. Although effective, these processing chains are computationally intensive and time-consuming. At the same time, the steady increase of onboard computational resources has encouraged a paradigm shift toward embedding decision-making capabilities directly within the satellite platform [56]. Tasks such as cloud detection and masking [4], image classification [40], anomaly detection [23], and object detection [14, 22] are increasingly being deployed onboard to reduce downlink bandwidth and latency. Onboard models do not benefit from the fully corrected imagery available in ground- based processing pipelines. A common strategy is to reproduce part of the restoration work- flow onboard, either through simplified physical corrections or lightweight neural approxi- mations [15, 41]. However, such approaches introduce additional computational overhead and partially defeat the purpose of moving intelligence to the edge. An alternative direction is to directly exploit raw satellite imagery as model input, thereby eliminating preprocessing stages. This strategy remains marginal, because raw data are scarcely available [41], and per- formance degradation is commonly observed when learning from uncorrected data [10, 45]. In particular, previous studies have shown a noticeable decline in object detection accuracy when operating on raw images instead of processed ones [14, 15], as convolutional architec- tures struggle to capture inter-class variations under degraded conditions. Object detection nevertheless remains a central task in remote sensing applications [42]. Convolutional detectors have long dominated the field, while more recent transformer-based architectures have demonstrated superior representational capabilities in computer vision. Yet, attention mechanisms introduce significant computational challenges for onboard de- ployment. Softmax operations, high-dimensional matrix multiplications, and quadratic com- plexity with respect to the token count make them difficult to integrate into constrained hard- ware environments [51]. In light of these considerations, we argue that the attention mechanism is efficient at finding discriminative representations from raw satellite imagery, provided it can be refor- mulated to meet strict deployment constraints. In this work, we introduce Aper-GATES (Aperture-Gated Attention Transformer for Embedded Systems), an efficient attention for- mulation compatible with FPGAs accelerators. Taking advantage of this new model, we also introduce TriCCOT (Tri-Part Convolutional Conformal Transformer). This novel ob- ject detection architecture combines convolutional priors, conformal prediction principles, and Aper-GATES. The proposed approach is designed to operate directly on raw imagery while preserving computational feasibility, thereby bridging the gap between detection per- formance and embedded deployment requirements. 2 Related Work Object detection in remote sensing Object detection in remote sensing has been exten- sively investigated over the past decade. Conventional detection frameworks include both convolutional two-stage architectures [21, 25, 48] and one-stage variants [20, 47, 50], often reinforced with multi-scale feature pyramids to better capture objects of varying sizes [35]. Transformer-based detectors such as DETR [8] make use of the attention mechanism [51] and achieve strong performance, albeit at a higher computational cost. While these methods focus on detection accuracy under standard imaging conditions, they generally assume ac- cess to preprocessed data and do not explicitly address the constraints imposed by degraded or raw onboard imagery. DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION3 Onboard deployments for remote sensing applications The deployment of deep learn- ing models directly onboard satellites has recently gained momentum as a means of reducing transmission overhead and enabling real-time responses [56]. Embedded inference has been demonstrated for cloud detection [4], anomaly detection [12], and scene classification [40]. Several studies have explored learning from raw or minimally restored images to bypass computationally demanding correction pipelines [41, 45]. However, studies report perfor- mance drops when models are trained and evaluated on degraded data [14], highlighting the difficulty of extracting robust representations without full radiometric and geometric pro- cessing. Rather than improving preprocessing stages, our approach tackles this limitation at the architectural level. Embedded object detection Although transformer-based vision models offer strong rep- resentation learning [8], their attention mechanisms remain difficult to deploy on resource- constrained onboard accelerators. Hybrid architectures such as CvT [53] and MT [24] incorporate convolutional inductive biases within attention-based models. Nevertheless, standard self-attention mechanisms [51] scale quadratically with token number and rely on softmax normalisation, while linear approximations [1] still require global matrix op- erations. Such characteristics limit their compatibility with resource-constrained onboard hardware. Existing FPGA deployments in remote sensing have therefore largely concen- trated on convolutional networks using platforms such as Xilinx Versal and DPU architec- tures [6, 19, 29, 32]. In contrast, transformer acceleration typically demands specialised datapaths or dedicated quantisation-aware schemes [30, 34, 54]. By reformulating attention without explicit token-token affinity computation and expressing it through convolutional projections combined with global channel statistics, our method enables integration within a standard CNN-oriented acceleration pipeline, reconciling transformer expressiveness with strict hardware constraints. Conformal prediction Conformal prediction has recently emerged as a principled tool for providing distribution-free uncertainty guarantees [3]. Extensions to object detection enable the construction of bounding boxes with probabilistic coverage guarantees [2, 9], primar- ily focusing on calibration and reliability. In contrast, we leverage conformal prediction as a structural component of the detection pipeline, systematically enlarging region propos- als before classification. This design ensures contextual completeness and mitigates biases introduced during proposal generation, without resorting to heuristic scaling rules. 3 Combining convolution, conformal and attention 3.1 Model overview TriCCOT comprises three sequential stages: a convolutional Region Proposal Network (RPN), conformal adjustment of the predicted boxes, and crop classification with Aper-GATES, our FPGA-friendly attention model. Processing object-centric crops reduces irrelevant image content and enables a compact classifier. The complete architecture is shown in Fig. 1. 4DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION Figure 1: TriCCOT architecture. It is composed of a Region Proposal Network, a conformal predictor, and Aper-GATES, our convolutional self-attention classifier. The RPN predicts bounding boxes that are then refined using a conformal predictor. Cropped images are cre- ated from the conformal bounding boxes, which are aggregated into a 2D token space that serves as input for the attention-based classifier. 3.2 Region Proposal Network (RPN) The core concept of the region proposal network derives from two-stage detection models [21] that rely on a first detector to predict bounding boxes and a second classifier to predict the class of the detected objects. Our RPN is based on a CSPDarknet backbone coupled with an FPN mechanism [35] to adapt to multiscale objects. As studies have shown that degradations introduced into raw data have little effect on bounding box localisation [14], few modifications are required to the RPN, and other convo- lutional backbones can be used if needed. 3.3 Conformal predictor RPN bounding boxes can be imperfect, leading to cropped or shifted objects that lack crucial information for classification. In addition, ground-truth annotations are typically tight around the object and may therefore provide little contextual information when directly converted into image crops. We therefore introduce a conformal predictor to relax the box constraints before feeding into the transformer. When applied to object detection, a conformal predictor aims to predict bounding boxes with probabilistic guarantees [2, 9]. For a bounding box Y = (Y x min ,Y y min ,Y x max ,Y y max ), the calibrated conformal predictor adjusts each predicted coordinate independently as b C α = n box with coordinates b Y j new â d j α/4 for jâx min , y min , x max , y max o (1) where d j α/4 is the calibrated correction associated with coordinate j. These corrections are estimated independently for each boundary and may therefore differ in both magnitude and sign, allowing the conformal predictor to compensate not only for insufficient box size but also for systematic localisation biases. Applying conformal prediction to the RPN outputs ensures that, under the conformal coverage assumptions, a fraction 1â α of ground-truth objects are entirely contained within the adjusted predicted bounding boxes. The resulting crops consequently provide additional contextual information to the attention-based classifier. Compared with a fixed multiplicative expansion ratio, conformal prediction provides a data-driven adjustment that accounts for the localisation errors and directional biases of the RPN with limited computational overhead. In this work, conformal prediction is applied by using the PUNCC library [38, 39]. DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION5 3.4 Aper-GATES: An Aperture-Gated Attention Transformer for Embedded Systems Transformer architectures have demonstrated strong performance across a wide range of tasks, primarily due to their ability to model long-range dependencies. Unlike convolutional neural networks, which are limited to the receptive fields of their kernels, Vision Trans- formers (ViTs) can aggregate features globally across an image. We believe that this global context can improve model robustness on degraded images. Figure 2: Aper-GATES architecture. It takes cropped images as input, maximising useful information. Cropped images are resized into pre-fixed sizes and projected into an embed- ding space. The attention matrix is computed using a gated convolutional attention module. Finally, a 2D convolutional mapping outputs the prediction. However, deploying transformers on edge hardware (particularly FPGAs) remains non- trivial due to the high cost of the softmax operation and the quadratic complexity of the attention mechanism. To address these constraints, we propose Aper-GATES (Aperture- Gated Attention Transformer for Embedded Systems). Inspired by Squeeze-and-Excitation and gated approaches [18, 28, 53], Aper-GATES utilises a hardware-efficient gating mecha- nism (analogous to an optical aperture) to modulate the flow of global information. Unlike SE-Net and GC-Net, which pool the feature map itself and reinject the resulting descriptor, Aper-GATES pools key-value interactions to obtain a global correlation descriptor that mul- tiplicatively gates the query tensor Q. Aper-GATES trades pairwise token expressivity for architectural simplicity, numerical stability, and compatibility with FPGA deployment, while preserving global context aggregation. The global architecture of Aper-GATES is displayed in Fig. 2. First, the crop goes through a dynamic resizing step. The resulting image is embed- ded in a latent space and processed by a gated convolutional attention mechanism. Finally, the attention result is processed by a convolutional layer that outputs the class probabilities. 3.4.1 Multi-resolution patch embedding bank with fixed grid Transformers can adapt to various input sizes. In the field of Large Language Models (LLMs), self-attention operates over variable-length token sequences in a latent space [51, 55]. Conversely, ViTs are traditionally trained with a fixed input resolution (usually 224x224) [8, 16, 26]. Nevertheless, some studies have leveraged this property to support variable spa- tial scales, such as FlexiViT, Swin Transformer, or DINOv2 [5, 37, 43]. In practice, varying the image resolution generally changes the number of input tokens or requires positional- embedding interpolation, which is undesirable for static FPGA deployment. Since Aper- GATES takes cropped images of variable sizes from the conformal predictor as input, it is important to retain the size and shape information of the cropped objects. However, support- ing variable input sizes can degrade FPGA performance. Therefore, we introduce a multi- resolution patch embedding that resizes the cropped images into fixed shapes, as displayed in Fig. 3. 6DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION Figure 3: Multi-resolution patch embedding pipeline. The size selector is implemented using Boolean conditioning on the CPU. A fixed-size embedding space is achieved by carefully selecting the convolutional kernel for each admissible input resolution, which improves DPU deployment by keeping the data shape fixed throughout the model. Finally, the metadata are encoded into the embedding space. Let the input crop be denoted by xâR BĂ3ĂHĂW , where the spatial size (H,W) is selected from a predefined set of admissible resolutions and is closest to the original crop size while preserving aspect ratio. In our implementation H,W â16, 32, 64, 128. Rather than using a single patch embedding layer, we define a bank of convolutional patch embedders, each specialised to a single admissible input size. For an input of size (H,W), the corresponding branch is selected. Because this selection relies on Boolean conditions, it is outsourced to the CPU rather than executed on the FPGA DPU. Since this operation remains fast, it does not significantly degrade inference speed. Each branch maps the input crop to a fixed spatial grid of size GĂ G, with G = 8 in our implementation, allowing a single model to be used for all admissible input sizes. This is achieved via two strided convolutional stages: z = Embed H,W (x)âR BĂDĂGĂG .(2) where the total downsampling stride is s h = H G ,s w = W G (3) For example, strides 2, 4, 8, 16 are split as (1, 2), (2, 2), (2, 4), and (4, 4), respectively. This allows all supported input sizes to be projected to the same 8Ă 8 feature grid. To provide the encoder with explicit knowledge of the input geometry, we compute a normalised metadata vector m = [g branch , m size ], where g branch is the one-hot encoded branch selection vector and m size is a normalised size vector computed as: m size = H H max , W W max , H/G P h,max , W/G P w,max .(4) where P h,max and P w,max denote the maximum patch-grid height and width allowed by the size selector, respectively. This metadata is projected via a two-layer convolutional network and integrated into the feature map z via additive residual fusion: z âČ = z+ CONV(m). This ensures that, while the feature grid is fixed, the network remains aware of the original object scale. DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION7 3.4.2 Gated Convolutional Attention Let XâR BĂCĂHĂW be an intermediate feature map, where B denotes the batch size, C the number of channels, and HĂ W the spatial resolution. We use the multihead strategy to obtain DPU-compatible batch sizes by partitioning the channels into h heads of dimension d, where C = hd. Figure 4: Aper-GATES attention mechanism. Context is gathered by a global descriptor. While convolutional projections for Q, K, V have been explored in prior works such as CvT [53] and CMT [24] to intro- duce spatial inductive bias, these architec- tures typically construct tokenâtoken affin- ity matrices that retain quadratic attention complexityO(N 2 ). In contrast, the core of Aper-GATES is a Gated Convolutional At- tention that trades pairwise interactions for a global second-order statistic, as displayed in Fig. 4, significantly reducing the compu- tational footprint. Convolutional Projections Instead of flattening spatial dimensions into N = HW tokens, we preserve the two-dimensional structure and compute head-wise projections using grouped 1Ă 1 convolutions: Q = Conv (g=h) 1Ă1 (X),K = Conv (g=h) 1Ă1 (X),V = Conv (g=h) 1Ă1 (X),(5) with Q, K, VâR BĂCĂHĂW . This operation maintains the spatial grid while projecting the input into the query, key, and value manifolds within each headâs subspace. Global Second-Order Context Rather than calculating an NĂ N affinity matrix, we take inspiration from the Squeeze-and-Excitation (SE) [7, 28] and Global Context (GC) [52] Net- works, and âsqueezeâ the spatial interaction into a global channel descriptor GâR BĂCĂ1Ă1 via global average pooling (P ) of the Key-Value product: G b,c =P(Kâ V) b,c = 1 HW H â u=1 W â v=1 (Kâ V) b,c,u,v (6) whereâ denotes the Hadamard product. By pooling the element-wise product of K and V, the model captures a global summary of feature correlations rather than localising specific token-to-token dependencies. Aperture Modulation Following the aperture metaphor, a gating function modulates the flow of this global context. The queries Q are âexcitedâ via multiplicative modulation with the transformed global descriptor: U = Qâ Ï(Proj(G)),(7) where Ï(·) represents a hardware-friendly activation (e.g., Hardsigmoid). This formulation allows the model to prioritise relevant global features while suppressing noise, thereby dy- namically gating spatial information. This mechanism achieves linear complexityO(BCHW), making it well-suited for resource-constrained deployment. 8DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION ConSmax Activation for DPU Deployment To avoid hardware-intensive softmax nor- malisation, Aper-GATES adopts a lightweight learnable gating function inspired by ConS- max [36]. Traditional ConSmax replaces softmax with a parametric transformation: ConSmax(x) = exp(xâ ÎČ) Îł ,(8) where ÎČ and Îł are learnable head-wise parameters. In our implementation, the exponential is replaced by a simplified DPU-compatible vari- ant based on depthwise affine transformation and a piecewise-linear Hardsigmoid: ConSmax DPU (x) = Hardsigmoid(w· x+ b).(9) The parameters w and b are learned channel-wise through a depthwise 1Ă 1 convolution. 3.5 Training and inference TriCCOT is a combination of three independent models that predict sequentially. There- fore, training is performed sequentially, with each model learning from the previous modelâs predictions. The RPN is first fitted on the detection training set; its outputs on a separate cal- ibration set are then used to fit the conformal corrections [3]. The conformal predictor guar- antees object integrity while avoiding overly large crops that would degrade inference time. In addition, because the conformal predictor is governed solely by Eq.1 during inference, it is both lightweight and computationally inexpensive to implement. Finally, Aper-GATES is trained from crops generated by the frozen RPN and conformal stages, using the same train/validation split as the RPN. The hyperparameters used to train TriCCOT are available in Tab. 1. Table 1: Training and model parameters for the DIOR and VDVRaw datasets. Param. CategoryParameterDIORVDVRaw RPNinput size640640 max detections7070 Conformal predictor α0.150.15 tested α[0.10, 0.20]- Aper-GATESpatch size16, 32, 64, 12832 embed. dim.128512 Trainingepochs200200 lr10 â3 10 â3 schedulerwarmup. cosinewarm. cosine Datasetnb. classes2017 train size18 0003 813 test size2 000425 3.6 Datasets The DIOR dataset [33] is chosen to evaluate the performance of TriCCOT, a widely used Earth-observation benchmark comprising 23,463 optical remote-sensing images across 20 DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION9 object classes. To assess robustness to degraded imagery, we also generate a degraded ver- sion of DIOR by applying spatial blur and signal-dependent noise that are representative of raw satellite imaging conditions. The degradation process approximately reverses restoration operations typically applied to satellite imagery [14]. In addition, we evaluated TriCCOT on real raw satellite imagery with the VENÎŒS raw images for vessel detection (VDVRaw) dataset [11]. VDVRaw contains 282 multispectral images divided into 4 238 patches affected by various sensor and acquisition artefacts, in- cluding striping noise, band misalignment, radiometric noise, blur, and stray light. It con- tains a total of 3 827 annotated objects across 17 classes. 3.6.1 Degradation simulation The degraded DIOR dataset combines spatial and radiometric degradations representative of realistic imaging systems. Spatial degradation via MTF Spatial resolution loss is modelled using a parametric Mod- ulation Transfer Function (MTF) combining optical blur and sensor sampling: MTF( f x , f y ) = e âÎł f r sinc( f x )sinc( f y ),f r = q f 2 x + f 2 y ,(10) where Îł is calibrated from a user-defined Nyquist value MT F Nyq , Îł =â2 log MTF Nyq sinc(0.5) .(11) The corresponding point-spread function is obtained by inverse Fourier transform and con- volved independently with each image channel. In our work, MT F @Nyquistâ [1%, 2%]. Signal-dependent noise Radiometric degradation is modelled using signal-dependent Gaus- sian noise, Ï 2 (L) = α L+ ÎČ,(12) where L denotes luminance. The parameters α and ÎČ are determined from two reference luminanceâSNR pairs using Ï 2 i = L i 10 SNR i,dB /20 2 = α L i + ÎČ.(13) Noise with standard deviation Ï(L) = p max(0, α L+ ÎČ)(14) is then sampled independently for each pixel. In our work, the SNR values of the simulated images for two points of luminance L 0 = 25W/m 2 /sr/ÎŒm and L 1 = 100W/m 2 /sr/ÎŒm are set between [25dB, 35dB] and [65dB, 75dB], respectively. 10DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION Table 2: Ablation study of the TriCCOT model on the DIOR dataset. ConfigurationmAP@50mAP@75mAP@50:95PrecisionRecall Full TriCCOT76.49%40.08%45.51%76.21%53.79% w/o Aper-GATES75.67%39.94%44.09%70.19%38.99% w/o ConSmax74.30%39.09%43.18%68.97%25.09% w/o Multi Res. Patch Embedding72.55%38.07%41.55%63.54%20.76% w/o Conformal Pred.69.10%38.60%42.67%65.21%25.51% 4 Results 4.1 Ablation study To ensure that all components of TriCCOT contribute to detection, we performed an abla- tion study on the original DIOR dataset. Aper-GATES is replaced with a traditional ViT implementation, ConSmax is replaced with a plain Hardsigmoid without learnable w, b, the multi-resolution patch embedding is replaced with a fixed 128x128 resize, and we com- pletely removed the conformal predictor. The results of the ablation study are available in Tab. 2. First, we observe that the full implementation of TriCCOT has the best performance on all criteria. This shows that each module contributes to improving detection quality. Remov- ing Aper-GATES results in only a modest reduction in performance, with a decrease of only 0.82 percentage points in mAP@50. This is because Aper-GATES does not aim to improve the attention mechanism, but rather to enable its implementation on accelerated embedded hardware. Moreover, we can see that removing the conformal predictor has a significant im- pact on performance, with a 7.39-percentage-point decrease in mAP@50. This result shows that the context provided by enlarged bounding boxes helps the transformer process the crop, particularly for misaligned predictions, as demonstrated by the gap in mAP@50. 4.2 Detection performance on the DIOR and VDVRaw datasets TriCCOT is evaluated on both the original DIOR dataset and the raw simulated DIOR dataset. In addition, we evaluated a two-stage detector with Faster R-CNN [48], lightweight one-stage detectors with YOLOX-S [20] and NanoDet-Plus [46], and a real-time transformer- based detector with RT-DETR [57]. This comparison covers a broad range of architectures. Tab. 4 reports the results for all models on the three datasets. On the original DIOR dataset, RT-DETR achieves the best overall detection performance. Among FPGA-compatible models, YOLOX-S achieves the best mAP@50, with 78.7%, while NanoDet-Plus achieves the best mAP@50:95, with 47.9%. Nevertheless, TriCCOT re- mains close to these results, with a maximum difference of less than 2.5 percentage points for both mAP@50 and mAP@50:95, while requiring only 3.5M parameters and 8.2 GFLOPs. In contrast, DETR achieves 49.9% mAP@50 and 28.7% mAP@50:95, highlighting the higher training and deployment costs of a standard transformer detector in this setting, as it must learn spatial relationships [16]. To examine the behaviour of TriCCOT beyond aggregate metrics, Tab. 3 reports AP@50 for representative DIOR categories spanning different object types and performance levels. Performance ranges from 43.09% for Bridge to 91.92% for Airplane, indicating that TriC- COT achieves competitive detections across diverse categories rather than relying on a few DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION11 Table 3: AP@50 on representative DIOR categories. ClassAirplane Baseball Bridge Harbor Ship Storage Vehicle Windmill AP@5091.9291.7043.0966.3884.1373.0755.7084.34 Table 4: Benchmark comparison on DIOR Original, DIOR Raw, and VDVRaw datasets. Training times correspond to a single NVIDIA RTX 4090. Bold values indicate the best result among FPGA-compatible models, while underlined values indicate the best overall result. Model CharacteristicsDIOR Original (D l1 )DIOR Raw (D raw )VDVRaw MethodParamsGFLOPsTrain.mAP@50mAP@50:95mAP@50mAP@50:95mAP@50mAP@50:95 Non-FPGA-compatible transformer models DETR [8]41.5 M38.4 âŒ30 h49.928.749.127.78.65.6 RT-DETR [57]20.0 M60.0 âŒ30 h82.151.380.348.229.726.5 FPGA-compatible models Faster R-CNN [48]40.0 M89.0 âŒ8 h74.847.469.240.418.212.2 YOLOX-S [20]8.9 M26.84â5 h78.746.172.640.120.014.3 NanoDet-Plus [46]2.4 M1.84â5 h75.347.970.041.121.216.3 TriCCOT (ours)3.5 M8.24â5 h 76.545.574.842.225.318.0 high-performing classes. Evaluation on the raw DIOR dataset shows that RT-DETR still achieves the best overall results. However, among FPGA-compatible models, TriCCOT obtains the best performance on both metrics, with 74.8% mAP@50 and 42.2% mAP@50:95. Notably, the transformer- based architectures (TriCCOT, RT-DETR and DETR) experience the smallest decline in per- formance when evaluated on the degraded rather than the original dataset, supporting the hypothesis that attention-based mechanisms improve robustness to image degradations. Finally, evaluation on VDVRaw confirms this trend on real raw imagery. Although RT- DETR remains the best-performing model overall, TriCCOT achieves the best results among FPGA-compatible models, reaching 25.3% mAP@50 and 18.0% mAP@50:95. These re- sults show that TriCCOT provides a favourable trade-off between robustness to raw-image degradations and embedded deployability. Furthermore, Fig. 5 shows examples of TriCCOT detections for both the original and raw datasets, with examples of the impact of the conformal predictor on the crops. We see that TriCCOT produces qualitatively accurate detections with accurate bounding boxes and class predictions. In these examples, the degradations have little impact on detection. The resulting crops also illustrate the additional context provided by the bounding boxes generated by the conformal predictor. 4.3 FPGA deployment Finally, TriCCOT is successfully deployed on a Xilinx Versal VCK190 FPGA. The Versal has gained traction in the space community, thanks to its radiation-tolerant specifications [17, 44], and state-of-the-art performance [6, 29]. In this work, we use a Deep Learning Processor Unit (DPU) [32], originally designed to accelerate CNN models. Therefore, unlike other implementations of transformers on FPGA boards [13, 30, 34, 54], we do not alter the FPGA architecture to fit the attention operations. By doing so, we remain flexible and can easily combine multiple architectures, such as CNNs and transformers. Thanks to the novel attention implementation described in Sec. 3.4, all convolutional and attentional operations are mapped to the DPU, with none offloaded to the CPU. 12DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION Pred. on originalsPred. on degradedConformal crops Figure 5: TriCCOT prediction on sampled original and degraded images. Tab. 5 compiles the results obtained on the Versal. Here, we focus on edge performance metrics, including inference speed and energy consumption. To better understand the overall inference time of TriCCOT, its processing pipeline is broken down into its main components. The RPN is the most time-consuming stage, accounting for 29.1 ms of the total inference time of 35.5 ms. In contrast, the Aper-GATES classification stage runs in only 3.9 ms on the DPU. Despite being executed on the CPU, the conformal prediction and size-selector stages introduce only limited overhead, with execution times of 2.4 ms and 0.1 ms, respec- tively. For comparison, we also deployed YOLOX-S on the DPU, while DETR inference was performed on the Versal CPU because the DPU does not support its transformer-based archi- tecture. Thanks to its lightweight architecture, TriCCOT is the fastest model on the Versal, achieving 28.2 FPS on 640x640 patches and enabling real-time processing with transformer architectures. 5 Discussion Robustness of TriCCOT under degraded imaging conditions against convolutional ap- proaches While TriCCOT does not achieve the highest detection accuracy on the original DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION13 Table 5: Versal VCK190 inference performance on a 640Ă 640 image. Latencies are aver- aged over 100 runs. Runner creation is performed only once at initialisation. MethodHardwarePrec.Size (MB)Latency (ms)âFPSâPower (W)âPxls/s/Wâ Runner creationCPUâ929.3 (± 20.1)â Image preprocessingCPUâ44.3 (± 5.2)â TriCCOT (ours)CPU+DPUint810.035.5 (± 3.4)28.225462 824 RPNDPUint81.929.1 (± 3.0)â ConformalCPUfloat322.92.4 (± 0.2)â Size selectorCPUâ0.1 (± 0.0)â Aper-GATESDPUint85.23.9 (± 2.5)â DETR [8]CPUint8162.713 160 (± 400)0.0822.51 456 YOLOX-S [20]DPUint89.274.1 (± 3.6)13.525221 184 DIOR dataset, its main objective is not to maximise performance under ideal image condi- tions, but to preserve detection quality when onboard imagery is degraded by spatial blur and signal-dependent noise. In this setting, TriCCOT achieves the best performance among the evaluated convolutional methods. In addition, transformer architectures exhibited smaller performance drops between the original and degraded datasets. This supports our hypothesis that combining self-attention and global features improves the modelâs overall robustness to image degradations. Contribution of the conformal prediction stage The ablation study highlights the impor- tance of the conformal predictor. Extending the crops improves object coverage and adds contextual information. It also helps remove labelling biases in the dataset. The advantage of using conformal prediction is that it removes the need for arbitrary scaling parameters, compensates for model prediction biases, and provides a mathematically defined coverage guarantee. Role of Aper-GATES in embedded transformer deployment A key contribution of TriC- COT is its improved deployability of transformer-based models on FPGA. Aper-GATESâs convolutional reformulation of self-attention enables fast inference. In addition, it is possible to combine both convolutional and attention-based architectures within a single DPU. This design greatly simplifies TriCCOT deployment on the Versal VCK190 and enables simulta- neous execution of multiple model architectures without reconfiguring the DPU. Although TriCCOT does not match the performance of large foundation transformer models, it offers a practical trade-off between predictive performance and deployability in embedded systems. TriCCOT classifier Training a transformer is not trivial, as it typically requires large training datasets to mitigate the lack of local biases. Because the transformer classifier in TriCCOT processes only object crops, most of the input consists of task-relevant informa- tion. In our experiments, training Aper-GATES was straightforward and yielded fast conver- gence. All of TriCCOTâs parts (RPN, conformal predictor and Aper-GATES) can therefore be trained in only 4-5 h on a single NVIDIA RTX 4090, compared with approximately 30 h for the transformer baselines considered in our experiments. This reduced training cost is enabled by the smaller feature space and compact architecture. By combining convolutional processing for full-image search with transformer-based classification on regions of interest, TriCCOT supports compact and efficient transformer architectures. 14DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION Limitations of sequential training Despite these advantages, TriCCOT introduces ad- ditional training complexity. The architecture comprises three components: the RPN, the conformal predictor, and the Aper-GATES classifier. Each module is trained after the pre- vious one, which creates dependencies between stages. As a consequence, biases or errors introduced by the RPN may propagate to the conformal predictor and then to the classifier. Although the conformal stage mitigates incomplete or overly tight proposals, the current se- quential training strategy does not allow the full pipeline to be optimised end-to-end. Future work should therefore investigate joint or cooperative training strategies or hard-example mining that allow the proposal and classification stages to adapt to one another during train- ing. Overall, TriCCOT is competitive with state-of-the-art object detection models on original images, and performs best on noisy images against models deployable on FPGAs. It opens the possibility of a lighter transformer architecture with increased deployment capability on edge devices, specifically for onboard remote sensing applications. 6 Conclusion In this work, we introduced TriCCOT, a novel tri-part architecture that combines convolu- tion, conformal prediction, and self-attention for onboard object detection under degraded imaging conditions. The method is based on a convolutional region proposal network, a conformal prediction stage, and Aper-GATES, a hardware-friendly attention-based classi- fier. This design exploits the localisation efficiency of convolutional detectors, the proba- bilistic coverage guarantees of conformal prediction, and the robustness of attention-based representations, while remaining compatible with embedded FPGA deployment. Experiments on the DIOR dataset show that TriCCOT achieves competitive detection performance on clean imagery and outperforms the evaluated convolutional and FPGA- compatible baselines under raw imaging degradations. These results demonstrate that com- bining convolutional proposal generation with attention-based crop classification improves robustness to spatial blur and signal-dependent noise. Aper-GATES makes TriCCOT fully compatible with the FPGA DPU accelerators by replacing standard attention operations with convolutional projections and hardware-friendly gating. This enables CNN-Transformer in- ference without modifying the accelerator architecture. Despite these promising results, several limitations remain. TriCCOT is currently trained sequentially, which may propagate errors between the RPN, conformal predictor, and clas- sifier. Future work will investigate joint or cooperative training strategies through hard- example mining [31, 49] to better optimise the full detection pipeline. We will also investigate adapting the RPN to rotated bounding-box tasks. Indeed, exper- iments with remote sensing imagery show that similar objects can exhibit different rotations (such as tennis courts). While the size information is preserved, the ratio can vary greatly. Using rotated bounding boxes could benefit the classifier by harmonising bounding-box ra- tios across the same classes. Finally, the proposed region proposal network is based on a CSPDarknet backbone. This sets a hard limit on the total number of parameters and the total size of our model. Ongo- ing work focuses on replacing CSPDarknet with a lighter convolutional backbone, such as MobileNet [27], thereby drastically reducing TriCCOTâs overall size and greatly improving inference speed. DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION15 References [1] Kwangjun Ahn, Xiang Cheng, Minhak Song, Chulhee Yun, Ali Jadbabaie, and Suvrit Sra. Linear attention is (maybe) all you need (to understand transformer optimization), 2024. URL https://arxiv.org/abs/2310.01082. [2] Leo Andeol, Thomas Fel, Florence de Grancey, and Luca Mossina. Confident ob- ject detection via conformal prediction and conformal risk control: an application to railway signaling. In Harris Papadopoulos, Khuong An Nguyen, Henrik Boström, and Lars Carlsson, editors, Proceedings of the Twelfth Symposium on Conformal and Probabilistic Prediction with Applications, volume 204 of Proceedings of Ma- chine Learning Research, pages 36â55. PMLR, 13â15 Sep 2023. URL https: //proceedings.mlr.press/v204/andeol23a.html. [3] Anastasios N. Angelopoulos and Stephen Bates. Conformal prediction: A gentle intro- duction. Found. Trends Mach. Learn., 16(4):494â591, March 2023. ISSN 1935-8237. doi: 10.1561/2200000101. URL https://doi.org/10.1561/2200000101. [4] Cesar Aybar, Gonzalo Mateo-GarcĂa, Giacomo Acciarini, VĂt R ÌuĆŸi Ë cka, Gabriele Me- oni, Nicolas LongĂ©pĂ©, and Luis GĂłmez-Chova. Onboard cloud detection and atmo- spheric correction with efficient deep learning models. IEEE Journal of Selected Top- ics in Applied Earth Observations and Remote Sensing, 17:19518â19529, 2024. doi: 10.1109/JSTARS.2024.3480520. [5] Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes, 2023. URL https://arxiv. org/abs/2212.08013. [6] Jacob Brown, Colton Yates, Jeffrey Goeders, and Michael Wirthlin. Rareplanes de- tection using yolov5 on the versal adaptive soc. In 2025 IEEE Aerospace Conference, pages 1â9, 2025. doi: 10.1109/AERO63441.2025.11068464. [7] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local net- works meet squeeze-excitation networks and beyond, 2019. URL https://arxiv. org/abs/1904.11492. [8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kir- illov, and Sergey Zagoruyko. End-to-end object detection with transformers, 2020. URL https://arxiv.org/abs/2005.12872. [9] Florence de Grancey, Jean-Luc Adam, Lucian Alecu, SĂ©bastien Gerchinovitz, Franck Mamalet, and David Vigouroux. Object Detection With Probabilistic Guarantees. In SAFECOMP 2022, LNCS 13415, SAFECOMP 2022, LNCS 13415, MĂŒnchen, Ger- many, September 2022. URL https://hal.science/hal-03769683. This preprint has not undergone peer review or any post-submission improvements or cor- rections. The Version of Record of this contribution will appear in SAFECOMP 2022, LNCS 13415 proceedings. [10] Roberto Del Prete, Gabriele Meoni, Manuel Salvoldi, Domenico Barretta, Maria Daniela Graziano, Nicolas LongĂ©pĂ©, and Alfredo Renga. Enhanced maritime 16DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION monitoring via onboard processing of raw multi-spectral imagery by deep learning. In IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pages 1713â1717, 2024. doi: 10.1109/IGARSS53475.2024.10641068. [11] Roberto Del Prete, Manuel Salvoldi, Domenico Barretta, Nicolas LongĂ©pĂ©, Gabriele Meoni, Arnon Karnieli, Maria Daniela Graziano, and Alfredo Renga. Enhancing mar- itime situational awareness through end-to-end onboard raw data analysis. IEEE Jour- nal of Selected Topics in Applied Earth Observations and Remote Sensing, 18:16997â 17018, 2025. doi: 10.1109/JSTARS.2025.3584999. [12] Lorenzo Diana, Jia Xu, and Luca Fanucci. Oil spill identification from sar images for low power embedded systems using cnn. Remote Sensing, 13(18), 2021. ISSN 2072- 4292. doi: 10.3390/rs13183606. URL https://w.mdpi.com/2072-4292/ 13/18/3606. [13] Peiyan Dong, Mengshu Sun, Alec Lu, Yanyue Xie, Kenneth Liu, Zhenglun Kong, Xin Meng, Zhengang Li, Xue Lin, Zhenman Fang, and Yanzhi Wang. Heatvit: Hardware- efficient adaptive token pruning for vision transformers, 2023.URL https:// arxiv.org/abs/2211.08110. [14] Adrien Dorise, Marjorie Bellizzi, Adrien Girard, Benjamin Francesconi, and StĂ©phane May. Explaining raw data complexity to improve satellite onboard processing. In 2025 European Data Handling and Data Processing Conference (EDHPC), pages 1â 8, 2025. [15] Adrien Dorise, Marjorie Bellizzi, and Omar Hlimi. Rethinking satellite image restora- tion for onboard ai: A lightweight learning-based approach, 2026. URL https: //arxiv.org/abs/2604.12807. [16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognition at scale, 2021. URL https://arxiv.org/abs/ 2010.11929. [17] Arnaud Dufour, JĂ©rĂŽme Carron, François Pierron, Matthieu Fongral, David Dangla, Guillaume Bascoul, Françoise Bezerra, Julien Mekki, Florence Malou, and Pierre Mail- lard. 7nm finfet technology heavy ion sel evaluation using xilinx versal as case study. In 2021 21th European Conference on Radiation and Its Effects on Components and Systems (RADECS), pages 1â6, 2021. doi: 10.1109/RADECS53308.2021.9954564. [18] StĂ©phane dâAscoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: improving vision transformers with soft convolutional inductive biases*. Journal of Statistical Mechanics: Theory and Experiment, 2022(11), 2022. ISSN 1742-5468. doi: 10.1088/1742-5468/ac9830. URL http://dx.doi.org/ 10.1088/1742-5468/ac9830. [19] Luis M. GarcĂ©s-SocarrĂĄs, Raudel Cuiman, Flor Ortiz, Juan A. VĂĄsquez-Peralvo, Jorge L. GonzĂĄlez-Rios, Mouhamad Chehaitly, Arkadii Kazanskii, Sahar Malmir, Amirhossein Nik, Jan Thoemel, Sumit Kumar, Marcele Kuhfuss, Swetha Varadajulu, Eva Lagunas, Juan C. M. Duncan, Jorge Querol, and Symeon Chatzinotas. Onboard DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION17 machine learning for satellite edge computing: The spaice project use case. In 2025 European Data Handling & Data Processing Conference (EDHPC), pages 1â8, 2025. [20] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021, 2021. URL https://arxiv.org/abs/2107.08430. [21] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierar- chies for accurate object detection and semantic segmentation, 2014. URL https: //arxiv.org/abs/1311.2524. [22] Thomas Goudemant, Benjamin Francesconi, Houssem Farhat, Lionel Daniel, Olivier Thiery, Erwann Kervennic, Adrien Girard, and Seif Mzoughi. DĂ©tection de navires embarquable Ă bord de satellites. In Actes de la 4Ăšme Conference on Artificial Intel- ligence for Defense (CAID 2022), Actes de la 4Ăšme Conference on Artificial Intelli- gence for Defense (CAID 2022), Rennes, France, November 2022. DGA MaĂźtrise de lâInformation. URL https://hal.science/hal-03881738. [23] Thomas Goudemant, Benjamin Francesconi, Michelle Aubrun, Erwann Kervennic, In- grid Grenet, Yves Bobichon, and Marjorie Bellizzi. Onboard anomaly detection for marine environmental protection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:7918â7931, 2024. doi: 10.1109/JSTARS.2024. 3382394. [24] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. Cmt: Convolutional neural networks meet vision transformers, 2022. URL https://arxiv.org/abs/2107.06263. [25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(9):1904â1916, 2015. doi: 10.1109/TPAMI. 2015.2389824. [26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr DollĂĄr, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. URL https://arxiv. org/abs/2111.06377. [27] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient con- volutional neural networks for mobile vision applications, 2017.URL https: //arxiv.org/abs/1704.04861. [28] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132â7141, 2018. doi: 10.1109/CVPR.2018.00745. [29] Younis Ibrahim, Li Chen, and Tian Haonan. Deep learning-based ship detection on fpgas. In 2022 14th International Conference on Computational Intelligence and Com- munication Networks (CICN), pages 454â459, 2022. doi: 10.1109/CICN56167.2022. 10008312. 18DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION [30] Ehsan Kabir, Jason D. Bakos, David Andrews, and Miaoqing Huang. A runtime- adaptive transformer neural network accelerator on fpgas.Microprocessors and Microsystems, 120:105223, February 2026.ISSN 0141-9331.doi: 10.1016/ j.micpro.2025.105223.URL http://dx.doi.org/10.1016/j.micpro. 2025.105223. [31] Aybora Koksal, Onder Tuzcuoglu, Kutalmis Gokalp Ince, Yoldas Ataseven, and A. Ay- din Alatan. Improved hard example mining approach for single shot object detectors, 2022. URL https://arxiv.org/abs/2202.13080. [32] Guoyu Li, Pengbo Zheng, Jian Weng, and Enshan Yang. Dpuv4e: High-throughput dpu architecture design for cnn on versal acap, 2025. URL https://arxiv.org/ abs/2506.11441. [33] Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, 159:296â307, January 2020. ISSN 0924-2716. doi: 10.1016/j.isprsjprs.2019.11.023. URL http://dx.doi.org/10.1016/j. isprsjprs.2019.11.023. [34] Zhengang Li, Mengshu Sun, Alec Lu, Haoyu Ma, Geng Yuan, Yanyue Xie, Hao Tang, Yanyu Li, Miriam Leeser, Zhangyang Wang, Xue Lin, and Zhenman Fang. Auto-vit-acc: An fpga-aware automatic acceleration framework for vision transformer with mixed-scheme quantization, 2022. URL https://arxiv.org/abs/2208. 05163. [35] Tsung-Yi Lin, Piotr DollĂĄr, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection, 2017. URL https:// arxiv.org/abs/1612.03144. [36] Shiwei Liu, Guanchen Tao, Yifei Zou, Derek Chow, Zichen Fan, Kauna Lei, Bangfei Pan, Dennis Sylvester, Gregory Kielian, and Mehdi Saligane. Consmax: Hardware- friendly alternative softmax with learnable parameters. In Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design, ICCAD â24, New York, NY, USA, 2025. Association for Computing Machinery. ISBN 9798400710773. doi: 10.1145/3676536.3676766. URL https://doi.org/10.1145/3676536. 3676766. [37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted win- dows, 2021. URL https://arxiv.org/abs/2103.14030. [38] Mouhcine Mendil, Luca Mossina, Marc Nabhan, and Kevin Pasini. Robust gas demand forecasting with conformal prediction. In Conformal and Probabilistic Prediction with Applications, pages 169â187. PMLR, 2022. [39] Mouhcine Mendil, Luca Mossina, and David Vigouroux. Puncc: a python library for predictive uncertainty calibration and conformalization. In Conformal and Probabilis- tic Prediction with Applications, pages 582â601. PMLR, 2023. DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION19 [40] Gabriele Meoni, Marcus MĂ€rtens, Dawa Derksen, Kenneth See, Toby Lightheart, An- thony SĂ©cher, Arnaud Martin, David Rijlaarsdam, Vincenzo Fanizza, and Dario Izzo. The ops-sat case: A data-centric competition for onboard satellite image classification. Astrodynamics, 8(4):507â528, 2024. [41] Gabriele Meoni, Roberto Del Prete, Federico Serva, Alix De Beusscher, Olivier Colin, and Nicolas LongĂ©pĂ©. Unlocking the use of raw multispectral earth observation im- agery for onboard artificial intelligence. IEEE Journal of Selected Topics in Ap- plied Earth Observations and Remote Sensing, 17:12521â12537, 2024. ISSN 2151- 1535. doi: 10.1109/jstars.2024.3418891. URL http://dx.doi.org/10.1109/ JSTARS.2024.3418891. [42] Janne MĂ€yrĂ€, Elina A. Virtanen, Ari-Pekka Jokinen, Joni Koskikala, Sakari VĂ€kevĂ€, and Jenni Attila.Mapping recreational marine traffic from sentinel-2 imagery using yolo object detection models.Remote Sensing of Environment, 326:114791, 2025.ISSN 0034-4257.doi:https://doi.org/10.1016/j.rse.2025. 114791.URL https://w.sciencedirect.com/science/article/ pii/S0034425725001956. [43] Maxime Oquab, TimothĂ©e Darcet, ThĂ©o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, HervĂ© Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. URL https: //arxiv.org/abs/2304.07193. [44] Noah Perryman, Christopher Wilson, and Alan George. Evaluation of xilinx versal ar- chitecture for next-gen edge computing in space. In 2023 IEEE Aerospace Conference, pages 1â11, 2023. doi: 10.1109/AERO55745.2023.10115906. [45] Roberto Del Prete, Manuel Salvoldi, Domenico Barretta, Nicolas LongĂ©pĂ©, Gabriele Meoni, Arnon Karnieli, Maria Daniela Graziano, and Alfredo Renga. Enhancing mar- itime situational awareness through end-to-end onboard raw data analysis, 2024. URL https://arxiv.org/abs/2411.03403. [46] RangiLyu. Nanodet-plus: Super fast and high accuracy lightweight anchor-free object detection model. https://github.com/RangiLyu/nanodet, 2021. [47] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2016. URL https://arxiv.org/abs/ 1506.02640. [48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real- time object detection with region proposal networks, 2016. URL https://arxiv. org/abs/1506.01497. [49] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining, 2016. URL https://arxiv.org/ abs/1604.03540. 20DORISE, BELLIZZI, COHEN, MAY: TRICCOT FOR ONBOARD OBJECT DETECTION [50] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection, 2019. URL https://arxiv.org/abs/1904.01355. [51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. URL https://arxiv.org/abs/1706.03762. [52] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks, 2018. URL https://arxiv.org/abs/1711.07971. [53] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers, 2021. URL https: //arxiv.org/abs/2103.15808. [54] Can Xiao, Jianyi Cheng, and Aaron Zhao. Refining datapath for microscaling vits, 2025. URL https://arxiv.org/abs/2505.22194. [55] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision- language tasks, 2021. URL https://arxiv.org/abs/2103.03493. [56] Bing Zhang, Yuanfeng Wu, Boya Zhao, Jocelyn Chanussot, Danfeng Hong, Jing Yao, and Lianru Gao. Progress and challenges in intelligent remote sensing satellite systems. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:1814â1822, 2022. doi: 10.1109/JSTARS.2022.3148139. [57] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection, 2023.