DP2PNet: Diffusion-Based Point-to-Polygon Conversion for Single-Point Supervised Oriented Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Annotating Rotated Bounding Boxes (RBBs) for oriented object detection is resource-intensive.
• Single-point supervision offers efficiency but lacks crucial size and orientation data, necessitating complex model priors.

Purpose of the Study:

• To propose DP2PNet, the first diffusion model framework for single-point supervised oriented object detection.
• To overcome limitations of existing methods by reducing reliance on complex priors and fixed refinement stages.

Main Methods:

• DP2PNet utilizes a multi-scale consistent noise generator, replacing manual priors with Gaussian noise.
• A Noise Cross-Constraint module employs multi-instance learning for optimal noise point selection.
• A Semantic Key Point Aggregator forms pseudo-RBBs from aggregated noise points via graph convolution and convex hulls.

Main Results:

• DP2PNet achieves competitive mAP50 scores of 53.82% on DOTA-v1.0 and 53.61% on DIOR-R.
• The framework demonstrates robustness to noise and strong cross-dataset generalization capabilities.
• Dynamic adjustment of refinement stages is supported without requiring retraining.

Conclusions:

• DP2PNet presents an effective and flexible approach to single-point supervised oriented object detection.
• The diffusion model framework significantly reduces annotation labor while maintaining high performance.
• This method offers a promising direction for cost-effective and robust object detection.