Most previous benchmarks (like MPII or COCO) focus on images with a few prominent people. HKBoot focuses specifically on crowded scenes . The major challenge it addresses is the "double-counting" problem or false positives, where a model might predict two keypoints for the same body part (e.g., two left shoulders) or mistake background noise for a joint. Finished Version 102 Upd: Arrival Of The Goddess
The paper proposes a method to distinguish between individuals in a crowd without relying on slow, two-stage "top-down" detection (detecting people first, then estimating pose). Instead, it uses a tagging mechanism to associate keypoints with specific individuals directly. Xml Marker 22 License Key Link [OFFICIAL]