TriCCOT: Tri-part Convolutional Conformal Transformer for Onboard Space Object Detection
Adrien Dorise, Marjorie Bellizzi, Julia Cohen, Stéphane May
cs.CV · cs.AI
Quick verdict
CNN RPN + conformal box enlargement + DPU-friendly attention; strong FPGA results on DIOR/VDVRaw, but conformal coverage unverified and metrics single-run.
Summary
The paper introduces TriCCOT, a three-stage onboard object detector for raw satellite imagery: a convolutional RPN (CSPDarknet + FPN) proposes boxes; a conformal prediction step (PUNCC library) enlarges them per-coordinate with a claimed distribution-free coverage guarantee; and Aper-GATES, a transformer classifier redesigned for DPU execution, classifies the crops. Aper-GATES replaces token-token softmax attention with grouped 1x1 convolutional projections, a global K-V product pooled descriptor that gates the queries (SE/GC-style), and a Hardsigmoid-based ConSmax variant; a multi-resolution patch-embedding bank maps variable crops to a fixed 8x8 grid. On DIOR, TriCCOT reaches 76.5% mAP@50 and 45.5% mAP@50:95 with 3.5M parameters and 8.2 GFLOPs — within 2.5 points of the best FPGA-compatible baselines (YOLOX-S, NanoDet-Plus) — and is the best FPGA-compatible model on degraded DIOR (74.8/42.2) and real raw VDVRaw imagery (25.3/18.0). Ablations attribute the largest gain (+7.39 mAP@50) to conformal enlargement; replacing Aper-GATES with a standard ViT costs only 0.82 points. The full int8 pipeline runs on a Xilinx Versal VCK190 at 35.5 ms (28.2 FPS, 25 W) without DPU modification. Key limitations: the conformal coverage guarantee is never empirically verified, all accuracy results are single-run point estimates without variance, calibration-set construction is underspecified, and no code release is stated.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- TriCCOT reaches 76.5% mAP@50 on DIOR with 3.5M parameters and 8.2 GFLOPs, within 2.5 points of the best FPGA-compatible baselines while running entirely on a stock DPU.
- On degraded DIOR (74.8/42.2) and real raw VDVRaw imagery (25.3/18.0 mAP@50/50:95), TriCCOT outperforms all evaluated FPGA-compatible detectors.
- Conformal enlargement of RPN boxes is the largest single contributor to accuracy (+7.39 mAP@50 in the Table 2 ablation) by adding context and compensating RPN localisation bias.
- Aper-GATES reformulates attention with convolutional projections, global channel statistics and Hardsigmoid gating so all operations map onto an unmodified Versal DPU, running the full int8 pipeline at 35.5 ms (28.2 FPS, 25 W).
- Attention-based architectures degrade less than convolutional ones under raw-image degradations (transformer drops of 0.8-1.8 mAP@50 vs 5.3-6.1 for CNN baselines on DIOR-Raw).
Strengths
Complete FPGA deployment with an unmodified DPU
Table 5 gives a full latency/power budget on the Versal VCK190: total 35.5 ms (±3.4) at 28.2 FPS and 25 W in int8, broken into RPN 29.1 ms, conformal 2.4 ms (CPU), size selector 0.1 ms, Aper-GATES 3.9 ms; YOLOX-S on the same DPU takes 74.1 ms and DETR 13,160 ms on CPU. Such end-to-end deployment evidence is rare and concretely supports the deployability claim.
Component-wise ablation cleanly isolates each module
Table 2 shows removing conformal prediction costs 7.39 mAP@50 (76.49 to 69.10), removing multi-resolution patch embedding costs 3.94, ConSmax 2.19, and swapping Aper-GATES for a standard ViT only 0.82 — honestly demonstrating Aper-GATES is accuracy-neutral and valuable for deployment, not accuracy.
Evaluation spans simulated and real raw imagery
Honest limitations and degradation protocol
Weaknesses
Conformal coverage guarantee asserted but never measured
All accuracy results are single-run point estimates
Aper-GATES novelty is integration, not a new attention family
No code release and non-standard evaluation split
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method paper (CNES / IRT Saint-Exupéry) targets onboard object detection on raw, uncorrected satellite imagery deployed on FPGA. TriCCOT chains three stages: a convolutional RPN (CSPDarknet + FPN, Sec. 3.2), a per-coordinate conformal correction of the proposed boxes (Sec. 3.3, Eq. 1, PUNCC library, α=0.15), and Aper-GATES, a transformer crop classifier redesigned so every operation maps onto a stock Versal DPU (Sec. 3.4). The authors claim: (1) competitive accuracy on DIOR (76.5% mAP@50) with 3.5M parameters and 8.2 GFLOPs, within 2.5 points of the best FPGA-compatible baselines; (2) best-in-class FPGA-compatible results under degradation (74.8/42.2 on DIOR-Raw; 25.3/18.0 on VDVRaw); (3) a distribution-free coverage guarantee from the conformal stage; and (4) full int8 deployment at 28.2 FPS / 25 W without DPU modification. Claims 1, 2 and 4 are well supported by Tables 4-5; claim 3 is the weakest link.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is TriCCOT and how does it work?
TriCCOT is a three-stage onboard object detector for satellite imagery: a convolutional region proposal network (CSPDarknet + FPN) finds candidate boxes, a conformal prediction step statistically enlarges each box so that under its assumptions about 85% of ground-truth objects stay inside (α=0.15), and a transformer classifier called Aper-GATES — built from DPU-friendly convolution and gating operations instead of softmax attention — classifies the resulting crops.
How fast is TriCCOT on FPGA hardware?
On a Xilinx Versal VCK190 the full int8 pipeline runs in 35.5 ms per 640×640 image (28.2 FPS) at about 25 W, with the RPN taking 29.1 ms and the Aper-GATES classifier only 3.9 ms, without modifying the standard DPU. A deployed YOLOX-S baseline needed 74.1 ms on the same DPU, and DETR ran on the CPU at roughly 13 seconds per image.
How much does the conformal prediction stage improve detection?
In the ablation study on DIOR, removing conformal box enlargement cuts mAP@50 from 76.49% to 69.10% — a 7.39-point drop, the largest of any component. The enlarged crops add contextual information and compensate for RPN localisation bias, replacing hand-tuned box-scaling rules with a calibrated, data-driven adjustment.
Is TriCCOT's code available?
The paper does not mention a public code release. The datasets (DIOR and VDVRaw) are publicly available, and the image-degradation simulator is fully specified by equations, but the TriCCOT implementation itself is not stated as open source.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 appear only as captions in the extraction; the visual content (architecture diagrams, qualitative detections) could not be inspected.
- Table 1's patch-size row renders as '{16, 32, 64, 128} 32' in the extracted layout, leaving the VDVRaw value ambiguous — a formatting artifact, not evidence of truncation.
- The supplied metadata abstract ends mid-sentence, but the complete sentence appears in the extracted body text; the paper text runs complete through reference [57].
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.