Hier R-CNN: Instance-Level Human Parts Detection and A New Benchmark

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Instance-level human part detection is crucial for analyzing human actions and attributes.
• Existing datasets lack the scale and annotation richness for comprehensive human part detection.

Purpose of the Study:

• To introduce COCO Human Parts, the first large-scale, richly annotated dataset for instance-level human part detection.
• To provide a strong baseline model, Hierarchical R-CNN, for human part detection and relationship prediction.

Main Methods:

• Developed COCO Human Parts based on COCO 2017, annotating human parts with bounding boxes, types (face, head, hand, foot), relationships, and fine-grained classifications.
• Proposed Hierarchical R-CNN, an extension of Mask R-CNN, for end-to-end instance-level human part detection and subordinate relationship prediction.

Main Results:

• The COCO Human Parts dataset includes 268,030 person instances across 66,808 images, with an average of 2.83 parts per instance.
• The Hierarchical R-CNN model demonstrates effective performance in detecting human parts and their relationships within complex scenes.

Conclusions:

• COCO Human Parts addresses the need for a large-scale dataset, facilitating advancements in gesture recognition, keypoint detection, and virtual reality applications.
• The Hierarchical R-CNN provides a robust baseline for human part detection, with publicly available code and dataset to encourage further research.