[Paper Review] Unified-IoU: For High-Quality Object Detection

Xichun Luo, Zhihao Cai+2 more|arXiv (Cornell University)|Aug 13, 2024

Industrial Vision Systems and Defect Detection5 citations

TL;DR

The paper introduces Unified-IoU (UIoU), a dynamic, focal-style IoU loss for bounding box regression that emphasizes high-quality predictions and balances convergence speed. It shows improvements on VOC2007 and COCO2017, with caveats on dense datasets like CityPersons unless paired with Focal-inv.

ABSTRACT

Object detection is an important part in the field of computer vision, and the effect of object detection is directly determined by the regression accuracy of the prediction box. As the key to model training, IoU (Intersection over Union) greatly shows the difference between the current prediction box and the Ground Truth box. Subsequent researchers have continuously added more considerations to IoU, such as center distance, aspect ratio, and so on. However, there is an upper limit to just refining the geometric differences; And there is a potential connection between the new consideration index and the IoU itself, and the direct addition or subtraction between the two may lead to the problem of "over-consideration". Based on this, we propose a new IoU loss function, called Unified-IoU (UIoU), which is more concerned with the weight assignment between different quality prediction boxes. Specifically, the loss function dynamically shifts the model's attention from low-quality prediction boxes to high-quality prediction boxes in a novel way to enhance the model's detection performance on high-precision or intensive datasets and achieve a balance in training speed. Our proposed method achieves better performance on multiple datasets, especially at a high IoU threshold, UIoU has a more significant improvement effect compared with other improved IoU losses. Our code is publicly available at: https://github.com/lxj-drifter/UIOU_files.

Motivation & Objective

• Motivate improving bounding box regression beyond traditional IoU-based losses by focusing training on high-quality predictions.
• Propose a dynamic weighting scheme (Focal Box) that scales bounding boxes to alter loss emphasis across training
• Incorporate a Focal Loss-inspired dual attention to further optimize weights across quality anchors
• Introduce UIoU as a unified loss function that permits easy comparison with existing IoU-based losses
• Demonstrate effectiveness across standard benchmarks (VOC2007, COCO2017) and analyze dense-case behavior (CityPersons)

Proposed method

• Introduce Focal Box by scaling prediction and GT boxes to alter IoU and loss weight without extra complex computations.
• Anneal bounding-box attention with a ratio hyperparameter that shifts emphasis from low-quality to high-quality boxes over training, using strategies (linear, cosine, fractional).
• Adopt a Focal Loss-inspired weighting scheme by using the confidence deficit (1 - confidence) to scale the IoU-based loss.
• Combine these components into Unified-IoU (UIoU), enabling easy switching among IoU baselines (GIoU, DIoU, CIoU, etc.) for comparison.
• Experiment with VOC2007, COCO2017, and CityPersons to validate improvements and analyze high-quality box performance.

Experimental results

Research questions

• RQ1 How can bounding-box regression loss be re-weighted dynamically to prioritize high-quality predictions without sacrificing convergence speed?
• RQ2 Does a Focal-Loss-inspired attention mechanism improve high-precision object detection when integrated with IoU-based losses?
• RQ3 Can a Unified-IoU loss outperform existing IoU-based losses (e.g., GIoU, CIoU, SIoU) on standard benchmarks, particularly at higher IoU thresholds?
• RQ4 How does UIoU behave on dense datasets, and can Focal-inv strategies mitigate potential drawbacks?

Key findings

• On VOC2007, UIoU variants improve high-IoU detection; UIoU(linear) achieves mAP50-75 of 62.95 with a +1.78% relative gain over CIoU baseline.
• UIoU(linear) achieves mAP50 of 69.8 and mAP75 of 63.3 on VOC2007, with a relative gain of +1.94% and +2.31% over CIoU for respective metrics.
• On COCO2017, UIoU shows modest but consistent gains: mAP50 up by 0.2%, mAP75 up by 0.8%, mAP95 up by 0.44%, and mAP50-95 up by 0.5% over CIoU for 300 epochs.
• UIoU results indicate better localization quality at higher IoU thresholds, with consistent improvements across multiple datasets.
• On CityPersons, standard UIoU degrades performance; applying Focal-inv (a reversed focus on easy examples) yields improvements in high-quality detection (e.g., AP90) compared to CIoU and other baselines.
• Ablation shows that dynamic ratio scheduling (ratio) and Focal-box concepts contribute to convergence speed and high-quality detection, with Focal-inv providing notable gains in dense scenarios.