Reproduce “DAMO-YOLO : A Report on Real-Time Object Detection Design” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xianzhe Xu, Yiqi Jiang, Weihua Chen, Yilun Huang, Yuan Zhang, Xiuyu Sun
DAMO-YOLO is a real-time object detection framework from Alibaba Group that establishes a new benchmark for accuracy-latency trade-offs among YOLO series models on COCO, leveraging advanced architectural designs and training strategies. It achieves superior performance across various scales on both T4 GPUs (e.g., DAMO-YOLO-S* reaches 47.7 mAP at 3.83ms) and x86-CPUs (DAMO-YOLO-Ns at 32.3 mAP at 4.08 ms).
Key insights
- Optimizing directly for the latency-MAP curve (actual inference time) is more relevant for industrial applications than solely relying on FLOPs-mAP.
- A design principle favoring a robust, computation-heavy neck combined with a minimalist ZeroHead can achieve better overall accuracy-latency trade-offs by reallocating computational resources.
- Reparameterization, flexible channel dimensions, and accelerated fusion mechanisms are effective in significantly improving the efficiency and performance of feature pyramid networks.
- Explicitly aligning classification and regression costs in dynamic label assignment (AlignOTA) is crucial for improving detection quality and robustness during training.
- Strategic two-stage knowledge distillation, particularly with a shallow detector head, effectively transfers knowledge and boosts model performance, especially for smaller models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.