Reproduce “Localization Distillation for Dense 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
Zhaohui Zheng, Rongguang Ye, Ping Wang, Dongwei Ren, Wangmeng Zuo, Qibin Hou, Ming-Ming Cheng
Localization Distillation (LD) improves dense object detectors by transferring precise localization knowledge from a teacher to a student model, leveraging a probability distribution representation for bounding box edges and identifying valuable localization regions. This approach consistently boosts the Average Precision of lightweight detectors by up to 2.1 points and achieves state-of-the-art results on COCO without adding inference overhead.
Key insights
- Representing bounding boxes as probability distributions effectively captures localization uncertainty, providing richer information for distillation than traditional regression targets.
- Logit mimicking, when specifically applied through Localization Distillation, can be more effective than feature imitation for transferring localization knowledge in object detection.
- Separately distilling localization knowledge, particularly from valuable localization regions, is crucial and more impactful than semantic knowledge distillation for improving object detectors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.