Reproduce “Beyond Hungarian: Match-Free Supervision for End-to-End Object Detection” 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
Shoumeng Qiu, Xinrun Li, Yang Long
A training scheme for DETR-based object detectors eliminates the Hungarian algorithm, replacing it with a differentiable correspondence learning mechanism. This approach results in over 50% faster training latency and improves overall detection accuracy by +0.7 AP, with a notable +4.2 AP_L gain for large objects on the COCO val2017 dataset.
Key insights
- Explicit bipartite matching is not essential for end-to-end object detection; attention-based mechanisms can autonomously learn soft query-to-target assignments through continuous optimization.
- Directly supervising the correspondence learning process with broadcast costs, alongside a sparse query loss, enables stable and effective training without discrete assignment heuristics.
- Replacing CPU-intensive, discrete operations with fully GPU-accelerated tensor operations significantly enhances training efficiency and hardware compatibility for DETR models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.