Reproduce “Extending Dataset Pruning to Object Detection: A Variance-based Approach” 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
Ryota Yagi
Yagi's research extends dataset pruning to object detection tasks through a variance-based approach, which resolves challenges in prediction attribution, scoring, and aggregation for complex visual data. The proposed Variance-based Prediction Score (VPS) consistently outperforms existing methods on PASCAL VOC and MS COCO, achieving up to a >1 percentage point improvement in mAP at 90% pruning ratios.
Key insights
- Sample informativeness in object detection is strongly indicated by the variance of model predictions (IoU or confidence) across training epochs, rather than just consistently 'easy' or 'hard' examples.
- Pruning strategies that prioritize examples exhibiting high variance in prediction scores (VPS) yield superior performance compared to methods focusing solely on low average scores (consistently hard examples), which often lead to performance degradation.
- The principled extension of dataset pruning, with its components for attribution, variance-based scoring, and statistical aggregation, successfully generalizes across different object detection architectures, such as Faster R-CNN and YOLOv5m.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.