Reproduce “Gradient Harmonized Single-stage Detector” 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
Buyu Li, Yu Liu, Xiaogang Wang
Researchers from The Chinese University of Hong Kong's Multimedia Laboratory introduced the Gradient Harmonizing Mechanism (GHM), an adaptive approach that dynamically adjusts the training process of single-stage object detectors by re-weighting examples based on their gradient density. This method, applied to classification and regression tasks, allows RetinaNet to achieve 41.6 mAP on the COCO test-dev set with a ResNeXt-101 backbone, outperforming previous state-of-the-art single-stage detectors and some two-stage models.
Key insights
- The inherent imbalances in dense object detection training can be summarized as a disharmonious distribution of gradient norms, where easy examples have small gradients, hard examples have larger gradients, and outliers can introduce instability.
- Dynamically assigning a harmonizing parameter based on gradient density allows the model to adaptively focus on informative examples while down-weighting over-represented easy examples and potentially noisy outliers, making training more stable and robust.
- Harmonizing gradients for bounding box regression is crucial for improving precise localization, especially at higher Intersection over Union (IoU) thresholds, by ensuring that examples with small but important errors maintain their influence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.