Reproduce “Points as Queries: Weakly Semi-supervised Object Detection by Points” on OpenResearch

About this paper

Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, Jian Sun

This research from MEGVII Technology and Fudan University introduces Point DETR, an object detection framework that effectively leverages point annotations within a weakly semi-supervised setting to reduce annotation costs. The approach establishes point annotations as a viable, information-rich middle ground, outperforming baselines and achieving up to a 2.0 AP improvement over adapted multi-level detectors on the COCO dataset with 20% fully labeled data.

Key insights

• Point annotations provide a cost-effective yet information-rich form of weak supervision, offering instance-level location and category details that are crucial for object detection and are superior to image-level labels.
• Integrating a dedicated point encoder and relative regression into a Transformer-based architecture (Point DETR) allows for the intrinsic and effective utilization of sparse point supervision for precise bounding box prediction.
• Positional and category embeddings within the point encoder are critical for accurate box regression, with positional embeddings providing location context and category embeddings offering class-specific prior knowledge.

Summary by alphaXiv.