[Paper Review] MPDIoU: A Loss for Efficient and Accurate Bounding Box Regression

Siliang Ma, Yong Xu|arXiv (Cornell University)|Jul 14, 2023

Advanced Neural Network Applications217 citations

TL;DR

The paper introduces MPDIoU, a new IoU-based loss for bounding box regression that incorporates minimum point distance to improve convergence speed and accuracy, and demonstrates its superiority on object detection, scene text spotting, and instance segmentation benchmarks.

ABSTRACT

Bounding box regression (BBR) has been widely used in object detection and instance segmentation, which is an important step in object localization. However, most of the existing loss functions for bounding box regression cannot be optimized when the predicted box has the same aspect ratio as the groundtruth box, but the width and height values are exactly different. In order to tackle the issues mentioned above, we fully explore the geometric features of horizontal rectangle and propose a novel bounding box similarity comparison metric MPDIoU based on minimum point distance, which contains all of the relevant factors considered in the existing loss functions, namely overlapping or non-overlapping area, central points distance, and deviation of width and height, while simplifying the calculation process. On this basis, we propose a bounding box regression loss function based on MPDIoU, called LMPDIoU . Experimental results show that the MPDIoU loss function is applied to state-of-the-art instance segmentation (e.g., YOLACT) and object detection (e.g., YOLOv7) model trained on PASCAL VOC, MS COCO, and IIIT5k outperforms existing loss functions.

Motivation & Objective

• Motivate the need for a more informative bounding box regression loss beyond traditional IoU and L-norm losses.
• Propose MPDIoU as an IoU-based similarity metric that includes minimum point-distance penalties.
• Develop a practical, easy-to-compute MPDIoU-based regression loss and show its integration into existing models.
• Demonstrate empirical gains of MPDIoU on diverse 2D vision tasks and standard datasets.

Proposed method

• Define MPDIoU as IoU minus penalties based on the squared distances between the top-left and bottom-right corners of the predicted and groundtruth boxes, normalized by image size.
• Provide algorithms to compute MPDIoU between axis-aligned rectangles and derive the MPDIoU loss as L_MPDIoU = 1 - MPDIoU for bounding box regression.
• Show that MPDIoU can distinguish cases where two boxes have the same IoU but different corner alignments, improving regression guidance.
• Replace baseline regression losses in YOLOv7 and YOLACT with L_MPDIoU to evaluate performance gains.
• Establish conversion formulas and factors (areas, centers, widths, heights) from corner coordinates to compute MPDIoU components.

Experimental results

Research questions

• RQ1 Can MPDIoU provide a more informative gradient signal than existing IoU-based losses when bounding boxes share aspect ratios but differ in width/height?
• RQ2 Does integrating MPDIoU into state-of-the-art detectors improve localization accuracy and convergence across object detection, instance segmentation, and scene text spotting benchmarks?
• RQ3 How does MPDIoU perform relative to GIoU, DIoU, CIoU, and EIoU on standard datasets like PASCAL VOC, MS COCO, IIIT5k, and MTHv2?
• RQ4 Is MPDIoU robust when boxes do not overlap, and does it encourage tighter, less redundant predictions?

Key findings

• MPDIoU-based regression improves accuracy over several IoU-based losses across YOLOv7 and YOLACT on VOC and COCO.
• MPDIoU demonstrates faster convergence and higher performance for object detection, character-level scene text spotting, and instance segmentation.
• Experiments on VOC, COCO, IIIT5k, and MTHv2 show MPDIoU yields favorable results compared with GIoU, DIoU, CIoU, and EIoU losses.
• The MPDIoU loss remains bounded and provides informative gradients even when IoU is zero, due to corner-distance penalties.
• Theorem 3.1 asserts that when aspect ratios match, a predicted box inside the groundtruth yields a lower L_MPDIoU than a box outside, favoring tighter localization.
• Empirical results indicate MPDIoU offers notable improvements on AP75 and related metrics across evaluated tasks.