[Paper Review] Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression

Zhaohui Zheng, Ping Wang+4 more|arXiv (Cornell University)|Nov 19, 2019

Advanced Neural Network Applications27 references876 citations

TL;DR

Introduces Distance-IoU (DIoU) and Complete IoU (CIoU) losses for bounding box regression, achieving faster convergence and better accuracy than IoU and GIoU, and shows DIoU-NMS improves object detection performance.

ABSTRACT

Bounding box regression is the crucial step in object detection. In existing methods, while $\ell_n$-norm loss is widely adopted for bounding box regression, it is not tailored to the evaluation metric, i.e., Intersection over Union (IoU). Recently, IoU loss and generalized IoU (GIoU) loss have been proposed to benefit the IoU metric, but still suffer from the problems of slow convergence and inaccurate regression. In this paper, we propose a Distance-IoU (DIoU) loss by incorporating the normalized distance between the predicted box and the target box, which converges much faster in training than IoU and GIoU losses. Furthermore, this paper summarizes three geometric factors in bounding box regression, \ie, overlap area, central point distance and aspect ratio, based on which a Complete IoU (CIoU) loss is proposed, thereby leading to faster convergence and better performance. By incorporating DIoU and CIoU losses into state-of-the-art object detection algorithms, e.g., YOLO v3, SSD and Faster RCNN, we achieve notable performance gains in terms of not only IoU metric but also GIoU metric. Moreover, DIoU can be easily adopted into non-maximum suppression (NMS) to act as the criterion, further boosting performance improvement. The source code and trained models are available at https://github.com/Zzh-tju/DIoU.

Motivation & Objective

• Motivate bounding box regression losses aligned with IoU-based evaluation metrics.
• Address slow convergence and inaccuracy of IoU and GIoU losses.
• Propose DIoU and CIoU losses incorporating distance and aspect-ratio factors.
• Demonstrate improved performance on YOLOv3, SSD, and Faster R-CNN across VOC and COCO datasets.
• Show DIoU can be integrated into NMS for better suppression.

Proposed method

• Define DIoU loss by adding a normalized center-point distance penalty to the IoU loss: L_DIoU = 1 - IoU + rho^2(b, bgt) / c^2.
• Extend to CIoU loss by including aspect-ratio consistency term v with adaptive weighting alpha.
• Derive gradients for DIoU and CIoU, with stable handling of small w,h in CIoU.
• Incorporate DIoU as the NMS criterion (DIoU-NMS) to suppress redundant boxes.
• Integrate DIoU and CIoU into YOLO v3, SSD, and Faster R-CNN and evaluate on VOC 2007+2012 and COCO 2017.
• Provide source code and trained models publicly.

Experimental results

Research questions

• RQ1 Can directly minimizing normalized center-point distance (DIoU) yield faster and more accurate bounding box regression than IoU-based losses?
• RQ2 Does incorporating aspect-ratio consistency (CIoU) further improve convergence and accuracy across detector types and datasets?
• RQ3 Is DIoU-NMS more robust than standard NMS in occlusion and crowded scenes?
• RQ4 Do DIoU/CIoU losses provide consistent gains across one-stage and two-stage detectors on VOC and COCO?

Key findings

• DIoU loss converges faster than IoU and GIoU in regression simulations and improves AP and AP75 on VOC with YOLO v3 and SSD.
• CIoU loss further improves regression by considering overlap, center distance, and aspect ratio, yielding larger gains than DIoU in several settings.
• DIoU-NMS provides additional improvements over standard NMS in multiple detectors and datasets.
• For YOLO v3 on VOC 07+12, CIoU achieves 5.67% AP and 8.95% AP75 improvements over IoU/GIoU baselines; DIoU achieves 3.29% AP and 6.02% AP75 improvements.
• SSD on VOC shows DIoU and CIoU outperform IoU and GIoU, with CIoU achieving up to 2.59% AP and 2.92% AP75 gains over GIoU.
• Faster R-CNN on COCO shows CIoU providing notable gains in AP, AP75, and large and medium object performance; DIoU also offers improvements, especially for small objects when paired with DIoU-NMS.
• DIoU-NMS consistently outperforms original NMS across thresholds in VOC/COCO experiments.