Reproduce “Partial Weakly-Supervised Oriented Object Detection” on OpenResearch

About this paper

Mingxin Liu, Peiyuan Zhang, Yuan Liu, Wei Zhang, Yue Zhou, Ning Liao, Ziyang Gong, Junwei Luo and 3 more

This research introduces Partial Weakly-Supervised Oriented Object Detection (PWOOD), a new framework that achieves competitive oriented object detection performance by leveraging a small portion of cheap weak annotations (e.g., horizontal boxes or single points) alongside abundant unlabeled data. The method significantly outperforms existing weakly-supervised algorithms and approaches semi-supervised methods that use more expensive rotated bounding box annotations.

Key insights

• Complex object attributes like orientation and scale can be effectively inferred from minimal weak supervision (e.g., horizontal boxes, single points) by designing specific learning strategies that exploit data transformations and spatial relationships.
• Dynamically adjusting pseudo-label filtering thresholds based on the teacher model's confidence distribution significantly improves pseudo-label quality and model robustness compared to static thresholding.
• Combining a small amount of weak annotations with extensive unlabeled data through a semi-supervised framework can bridge the performance gap between purely weak supervision and more expensive full or semi-supervision.

Summary by alphaXiv.