Source-linked AI summary
PIoU Loss: Towards Accurate Oriented Object Detection in Complex Environments
Zhiming Chen, Kean Chen, Weiyao Lin, John See, Hui Yu, Yan Ke, Cong Yang
TL;DR
Existing OBB detectors often use distance losses that inadequately capture global IoU, particularly for high-aspect-ratio objects. This paper introduces a continuous pixel-wise PIoU loss that jointly reflects angle and IoU, evaluates it across anchor-based and anchor-free frameworks, and reports improved OBB detection performance while introducing Retail50K for challenging environments.
Problem
Existing OBB regression losses mainly minimize angle error and loosely correlate with IoU, limiting sensitivity to high-aspect-ratio objects.
Method
PIoU Loss is a continuous, differentiable pixel-wise IoU-derived loss that jointly models the five OBB parameters and overlapping-pixel contributions.
Results
PIoU loss significantly improves OBB-detector accuracy, particularly for objects with high-aspect ratios, across anchor-based and anchor-free frameworks.
Takeaways & Limitations
Retail50K extends evaluation to oriented targets with high aspect ratios, heavy occlusions, and complex backgrounds.
Takeaways & Limitations
The PIoU loss costs 15-20% more time than three other compared loss functions, while HPIoU offers lower computational complexity with slightly reduced performance.
Abstract
from arXiv · showhide
Object detection using an oriented bounding box (OBB) can better target rotated objects by reducing the overlap with background areas. Existing OBB approaches are mostly built on horizontal bounding box detectors by introducing an additional angle dimension optimized by a distance loss. However, as the distance loss only minimizes the angle error of the OBB and that it loosely correlates to the IoU, it is insensitive to objects with high aspect ratios. Therefore, a novel loss, Pixels-IoU (PIoU) Loss, is formulated to exploit both the angle and IoU for accurate OBB regression. The PIoU loss is derived from IoU metric with a pixel-wise form, which is simple and suitable for both horizontal and oriented bounding box. To demonstrate its effectiveness, we evaluate the PIoU loss on both anchor-based and anchor-free frameworks. The experimental results show that PIoU loss can dramatically improve the performance of OBB detectors, particularly on objects with high aspect ratios and complex backgrounds. Besides, previous evaluation datasets did not include scenarios where the objects have high aspect ratios, hence a new dataset, Retail50K, is introduced to encourage the community to adapt OBB detectors for more complex environments.
1 Introduction
Oriented bounding boxes reduce background overlap for rotated objects, but existing distance-based regression losses inadequately reflect IoU, especially for high-aspect-ratio targets. The paper introduces PIoU Loss and Retail50K to improve OBB detection in challenging environments.
- Motivation: OBBs extend horizontal boxes with a rotation angle, compactly enclosing rotated and densely crowded objects.An OBB is represented by (cx, cy, w, h, θ), whereas a horizontal box uses (cx, cy, w, h).
- Problem: Existing OBB approaches commonly optimize the added angle dimension with distance losses rather than global IoU.This limits performance in complex, close-up environments and makes the loss insensitive to high-aspect-ratio targets.
- Approach: PIoU Loss jointly targets angle and IoU accuracy through a continuous, differentiable pixel-wise formulation for OBB regression.It accumulates contributions from interior overlapping pixels and is evaluated in both anchor-based and anchor-free frameworks.
- Dataset: Retail50K is introduced as a benchmark for oriented targets with high aspect ratios, heavy occlusions, and complex backgrounds.The dataset addresses challenges not represented in earlier evaluation settings.
- Results: Experiments report improved performance for PIoU-based OBB detectors in both anchor-based and anchor-free settings and across different datasets.The reported benefits include aerial imagery and the more challenging Retail50K environment.
2 Related Work
Prior work extends horizontal-box detectors and develops standard regression or IoU losses, but oriented boxes introduce an angle dimension that these approaches do not directly accommodate. PIoU is presented within this gap as a general concept applicable to OBB-based frameworks.
- Oriented Object Detection: Most oriented object detectors extend generic horizontal-box detectors by adding an angle dimension to represent object orientation.Examples include rotation-invariant SSD, rotated Faster R-CNN, RoI transformers, and generative OBB proposal models.
- Regression Losses: Common regression losses include MSE, MAE, Quantile, Huber, and Log-Cosh losses.These losses operate on prediction errors, but OBB descriptors contain an additional angle dimension.
- PIoU: PIoU is described as a general concept applicable to most OBB-based frameworks, matching predicted and ground-truth OBB pairs before computing the final loss.The matching process is illustrated for predicted and ground-truth OBB pairs.
- Regression Losses: Standard regression losses cannot be directly used for OBB descriptors because of the additional angle dimension.This motivates specialized treatment of oriented-box regression.
- IoU-Based Losses: IoU and GIoU losses were developed for horizontal bounding-box regression, including cases where polygons do not intersect.The passage indicates that these approaches become problematic when applied to oriented bounding boxes.
3 Pixels-IoU (PIoU) Loss
PIoU constructs a differentiable, pixel-wise approximation of OBB IoU and uses it to guide regression toward accurate overlap. The method computes pixel contributions from relative position and applies the resulting intersection and union statistics to OBB loss, while also supporting horizontal boxes.
- Pixel-wise IoU construction: PIoU approximates OBB areas and overlaps by counting or weighting interior pixels within the smallest horizontal box covering both boxes.The method computes intersection and union statistics from pixels in image space.
- Pixel-wise IoU construction: A pixel’s relative position to an OBB is determined from distances formed by the box center, the pixel, and their perpendicular intersection point.The construction uses triangle-side distances and binary inside/outside constraints.
- Differentiable approximation: PIoU replaces the discontinuous IoU calculation with a product of two kernels, producing a continuous and differentiable pixel contribution for backpropagation.The kernel uses an adjustable factor k to control pixel sensitivity.
- Differentiable approximation: The pixel-based intersection and union statistics are inherently sensitive to OBB rotation and size, linking the loss to both geometric factors.This sensitivity follows from using relative pixel position, including distance and angle, in the kernel.
- Loss optimization: The proposed loss is computed over positive predicted–ground-truth box pairs to maximize PIoU during regression.Positive pairs are defined using matched anchors, with matching based on IoU greater than 0.5.
- Loss optimization: PIoU also applies to horizontal bounding-box regression by setting the rotation angle θ to 0, and it theoretically retains gradients when boxes do not intersect.The paper reports experimental validation for horizontal regression and states that PIoU remains positive without intersection.
4 Retail50K Dataset
Retail50K is a retail-environment benchmark designed to evaluate OBB detection under complex backgrounds, occlusions, varied orientations, and predominantly high aspect ratios.
- Compared with established OBB datasets focused mainly on aerial imagery, Retail50K targets complex close-up retail environments.
- Retail50K contains 47,000 supermarket images annotated with layer edges of shelves, fridges, and displays.
- Its retail scenes contain tightly packed items with varied colours and textures, while price and sale tags occlude layer edges by a mean of around 37.5%.
- The dataset was annotated by five skilled annotators using labelling rules, peer reviews, and 165 tasks grouped by image metadata.
5 Experiments
Experiments evaluate PIoU across anchor-based and anchor-free OBB detectors, oriented and horizontal bounding boxes, and multiple datasets. PIoU improves detection performance, especially for challenging high-aspect-ratio objects, while HPIoU offers lower complexity and faster training.
- Experimental Setup: PIoU was evaluated with anchor-based RefineDet and anchor-free CenterNet across DOTA, HRSC2016, Retail50K, and PASCAL VOC.The evaluation also compares different backbones and state-of-the-art OBB detectors.
- Ablation Study: k = 10 was selected for PIoU because it achieved the best accuracy among the tested sensitivity factors.The study varied k = 5, 10, and 15 on DOTA using RefineDet.
- Ablation Study: Around 3.5%: PIoU improved oriented bounding-box detection on DOTA compared with L1, SmoothL1, and L2 losses.HPIoU remained comparable to PIoU while reducing computational complexity.
- Ablation Study: PIoU outperformed SmoothL1 and GIoU for horizontal bounding-box regression on PASCAL VOC, particularly at high IoU thresholds.The comparison supports PIoU's use for both horizontal and oriented bounding boxes.
- Benchmark Results: ∼7% improvement for RefineDet-OBB and ∼6% improvement for CenterNet-OBB were reported on Retail50K with PIoU loss.The gains were more obvious than on DOTA, suggesting particular usefulness for high aspect ratios and complex environments.
- Benchmark Results: 55 fps v.s 12 fps: CenterNet-OBB+PIoU retained competitive performance while using a smaller image size than R3Det-800.The comparison used 512×512 images for the proposed detector and 800×800 for R3Det-800.
- Qualitative Results: PIoU produced more robust and accurate visual detections than SmoothL1 on Retail50K, HRSC2016, and DOTA, especially Retail50K.HPIoU was slightly lower than PIoU by 0.87, 1.41, and 0.18 mAP on DOTA, Retail50K, and HRSC2016, respectively, but trained faster.
6 Conclusion
The paper concludes that PIoU uses both angle and IoU information for accurate OBB regression and improves OBB detection, particularly for high-aspect-ratio objects. It also introduces Retail50K to examine these challenges and validate PIoU-based detectors.
- Conclusion: PIoU exploits both angle and IoU information through a pixel-wise IoU-derived loss for OBB regression.The loss is described as simple and suitable for horizontal and oriented bounding boxes.
- Conclusion: PIoU significantly improves OBB detector accuracy, particularly on objects with high-aspect ratios.The conclusion reports evaluations on both anchor-based and anchor-free frameworks.
- Conclusion: Retail50K is introduced as a challenging dataset for exploring existing OBB-detector limitations and validating PIoU.The paper notes future extension of PIoU to 3D rotated object detection and preliminary KITTI results.