PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection

Abstract: Single point supervised oriented object detection has gained attention and made initial progress within the community. Diverse from those approaches relying on one-shot samples or powerful pretrained models (e.g. SAM), PointOBB has shown promise due to its prior-free feature. In this paper, we propose PointOBB-v2, a simpler, faster, and stronger method to generate pseudo rotated boxes from points without relying on any other prior. Specifically, we first generate a Class Probability Map (CPM) by training the network with non-uniform positive and negative sampling. We show that the CPM is able to learn the approximate object regions and their contours. Then, Principal Component Analysis (PCA) is applied to accurately estimate the orientation and the boundary of objects. By further incorporating a separation mechanism, we resolve the confusion caused by the overlapping on the CPM, enabling its operation in high-density scenarios. Extensive comparisons demonstrate that our method achieves a training speed 15.58x faster and an accuracy improvement of 11.60%/25.15%/21.19% on the DOTA-v1.0/v1.5/v2.0 datasets compared to the previous state-of-the-art, PointOBB. This significantly advances the cutting edge of single point supervised oriented detection in the modular track.

• The paper introduces a novel detection method that uses non-uniform sampling and PCA to generate pseudo-rotated bounding boxes from single-point annotations.
• It achieves significant improvements in mAP with over 15.58x speedup and enhanced precision on challenging datasets like DOTA.
• PointOBB-v2 eliminates the resource-intensive teacher-student structure, offering a simpler, faster, and more memory-efficient approach for oriented object detection.

Overview of "PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection"

The research paper presents PointOBB-v2, an advanced methodology aimed at enhancing the efficiency, speed, and accuracy of single-point supervised oriented object detection without relying on extensive computational resources. The proposed technique efficiently generates pseudo-rotated bounding boxes using only point annotations, significantly improving over previous methods, such as PointOBB, by integrating innovative strategies in the pseudo-label generation process, focusing on non-uniform sampling and Principal Component Analysis (PCA).

Methodological Contributions

The paper introduces a simplified pipeline characterized by its removal of the resource-intensive teacher-student structure. The core innovation of PointOBB-v2 lies in its ability to effectively utilize Class Probability Maps (CPM) for inferring object boundaries and orientations. The authors propose a sampling strategy based on probability distributions and employ PCA to accurately discern object contours and orientations. They also introduce a mechanism to alleviate confusion in scenarios with high object density by ensuring effective separation of connected CPMs.

Key advancements from the previous state-of-the-art, PointOBB, include:

• Improved Sampling Strategy: A novel non-uniform positive and negative sampling method improves the CPM’s ability to delineate object contours accurately.
• PCA Integration: PCA leverages the probability-weighted sampling grid, facilitating precise orientation estimation and robust boundary determination.
• Resource Efficiency: PointOBB-v2 dramatically reduces training time, achieving over 15.58 times speedup compared to its predecessor, with memory consumption approximately at 8GB, bypassing the need for restrictions on Region of Interest (RoI) proposals.

Numerical Results and Implications

The experimental evaluations conducted on the DOTA-v1.0, v1.5, and v2.0 datasets serve to underscore the efficiency of PointOBB-v2. The method achieves notable improvements in mean Average Precision (mAP), specifically recording relative enhancements of 11.60%, 25.15%, and 21.19% on these datasets respectively. The paper provides thorough dataset-specific analyses, showing significant gains especially in challenging scenarios populated with dense and small objects where previous methods struggled.

These results demonstrate the capability of PointOBB-v2 to deliver high precision in object detection tasks without traditional heavy reliance on large-scale annotations, making it particularly of interest for applications in remote sensing and various real-time detection systems where speedy and memory-efficient solutions are crucial.

Future Directions

The paper hints at several future research directions, particularly focused on further automating and refining the point annotation accuracy to fully harness the system’s potential. The authors also suggest exploring expansions of this model to incorporate other forms of weak supervision or constraints for additional performance boosts. Given its significant efficiency improvements, PointOBB-v2 paves the way for future-oriented object detection systems, potentially aligning with deployment scenarios across different domains outside aerial imagery where computational resources are at a premium.

The innovative sampling strategy and PCA application model the future of resource-efficient, point-supervised detection, a valuable step towards democratizing complex AI models for real-time applications.