Reproduce “Detector-in-Detector: Multi-Level Analysis for Human-Parts” 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
Xiaojie Li, Lu Yang, Qing Song, Fuqiang Zhou
A Detector-in-Detector Network (DID-Net) framework is introduced to enhance the detection of multi-level human parts, specifically body, face, and hand, by explicitly leveraging their hierarchical relationships. This approach achieves a mean Average Precision (mAP) of 91.1% on a newly created Human-Parts dataset, outperforming existing state-of-the-art methods for person and hand detection.
Key insights
- Explicitly modeling hierarchical relationships between objects (e.g., body and its parts) through a coarse-to-fine cascaded detection mechanism significantly improves the accuracy of small sub-object detection.
- Focusing a dedicated 'PartsDetector' on cropped parent object regions reduces background clutter and provides higher resolution features, which is crucial for overcoming challenges in small object detection.
- A two-stage RPN+RCNN design within the PartsDetector, leveraging multiple RoI Align operations, proved more effective for refining hand and face detection compared to simpler architectural choices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.