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

Botao Ren, Xue Yang+3 more|arXiv (Cornell University)|Oct 10, 2024

Advanced Neural Network Applications4 citations

TL;DR

PointOBB-v2 proposes a faster, simpler, and more accurate method for single point-supervised oriented object detection by generating pseudo rotated boxes via Class Probability Maps (CPM) and Principal Component Analysis (PCA), eliminating the need for a teacher-student structure. It achieves 15.58× faster pseudo-label generation and boosts mAP by 11.60% to 25.15% on DOTA-v1.0 to v2.0, with superior performance on small and densely packed objects.

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.

Motivation & Objective

• To address the inefficiency and high memory usage of existing point-supervised oriented detection methods, particularly the slow pseudo-label generation in PointOBB.
• To eliminate reliance on teacher-student frameworks and external priors, improving training speed and memory efficiency.
• To enhance detection accuracy in dense object scenarios, especially for small and overlapping objects.
• To develop a modular, scalable pipeline that decouples pseudo-label generation from detector training for greater flexibility.

Proposed method

• The method generates a Class Probability Map (CPM) using non-uniform positive and negative sampling to learn approximate object regions and contours.
• Principal Component Analysis (PCA) is applied to the CPM to estimate object orientation and bounding box boundaries with high accuracy.
• A non-uniform sampling strategy based on the CPM's probability distribution improves localization precision and orientation estimation.
• A separation mechanism is introduced to resolve confusion in pseudo-labels caused by overlapping CPMs in dense object regions.
• The approach discards the teacher-student structure used in PointOBB, significantly reducing training time and GPU memory consumption.
• A vector constraint module is applied to further improve performance on dense object categories like small vehicles and ships.

Experimental results

Research questions

• RQ1 Can a simpler, faster, and more accurate pseudo-label generation pipeline be designed for single point-supervised oriented object detection without relying on teacher-student frameworks or external models?
• RQ2 How can Class Probability Maps (CPM) be effectively leveraged to estimate object orientation and boundaries using PCA in a weakly supervised setting?
• RQ3 To what extent does the proposed method improve detection accuracy and efficiency in high-density and small-object scenarios compared to PointOBB?
• RQ4 Can the method maintain robustness under noisy point annotations, particularly when center points are imperfectly labeled?
• RQ5 How does the separation mechanism enhance pseudo-label quality in overlapping or densely packed object regions?

Key findings

• PointOBB-v2 achieves a 15.58× speedup in pseudo-label generation, reducing training time from 22.28 hours to 1.43 hours on DOTA-v1.5.
• On DOTA-v1.0, the method improves mAP from 30.08% (PointOBB) to 41.68%, a gain of 11.60%.
• On DOTA-v1.5, mAP increases from 30.65% to 36.39%, a 25.15% improvement, with significant gains in small and dense object categories.
• On DOTA-v2.0, mAP rises from 26.63% to 27.22%, a 21.19% relative improvement, demonstrating robustness in extreme density scenarios.
• The method maintains stable performance under point annotation noise, with only a 2.27% mAP drop under 10% center point perturbation, outperforming PointOBB.
• Visualizations confirm that PointOBB-v2 generates non-overlapping, accurate pseudo-labels for densely packed objects like small vehicles, while PointOBB produces overlapping and ambiguous predictions.