Source-linked AI summary
MPDIoU: A Loss for Efficient and Accurate Bounding Box Regression
Siliang Ma, Yong Xu
TL;DR
Existing bounding box regression losses can fail when predicted and groundtruth boxes share an aspect ratio but differ in width and height. The paper proposes MPDIoU, a minimum-point-distance similarity metric, and its LMPDIoU loss; experiments report improved performance across detection, scene text spotting, and instance segmentation benchmarks.
Problem
Most existing bounding box regression losses cannot be optimized when predicted and groundtruth boxes have the same aspect ratio but different width and height values.
Method
The paper proposes MPDIoU, which uses minimum point distances while incorporating overlap, center distance, and width-height deviation, and defines LMPDIoU from it.
Results
MPDIoU improves performance on object detection, scene text spotting, and instance segmentation benchmarks using commonly used measures and MPDIoU.
Takeaways & Limitations
MPDIoU is presented as a simpler similarity metric and a suitable IoU substitute for performance measures in 2D/3D vision tasks, while this paper applies it to 2D detection and segmentation.
Takeaways & Limitations
This paper focuses on 2D object detection and instance segmentation; extension to non-axis-aligned 3D cases is left for future work.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper motivates improved bounding box regression losses for object detection and instance segmentation, then proposes MPDIoU-based regression to address existing-loss limitations and evaluate it broadly.
- Bounding box regression is a core module in many object detectors and instance segmentation systems, making its loss function important for localization.
- The proposed LMPDIoU loss is designed from the strengths and weaknesses of existing IoU-based and ℓn-norm losses.The authors target faster convergence and more accurate regression results.
- Experiments evaluate the proposed loss on object detection, character-level scene text spotting, and instance segmentation tasks.The study uses datasets including PASCAL VOC, MS COCO, IIIT5k, and MTHv2.
- Extensive experiments and ablation studies report the superiority of MPDIoU and examine loss-function settings and parameter values.
2. Related Work
Prior bounding box regression losses evolved from norm-based objectives to IoU-based metrics incorporating overlap, enclosure, center distance, and aspect-ratio information. The paper identifies a shared failure case and motivates MPDIoU as a more efficient alternative based on minimum point distances.
- Bounding box regression losses progressed from scale-sensitive ℓn-norm objectives to IoU-based metrics used for training detection and segmentation models.
- GIoU adds an enclosing-box penalty so nonoverlapping predictions can move toward the target, but loses effectiveness when the prediction is fully covered by the groundtruth.
- DIoU incorporates the Euclidean distance between predicted and groundtruth box centers, while coincident centers reduce it to original IoU.
- CIoU adds aspect-ratio information, whereas EIoU addresses CIoU’s relative rather than absolute aspect-ratio definition.
- Existing losses can become ineffective when prediction and groundtruth share an aspect ratio but have different widths and heights.MPDIoU instead minimizes distances between corresponding top-left and bottom-right points while retaining factors from prior losses.
3. Intersection over Union with Minimum Points Distance
MPDIoU represents bounding-box similarity using minimum distances between corresponding corner points, while retaining area, center-distance, and size-deviation factors. Its loss supports overlapping and nonoverlapping regression and distinguishes same-aspect-ratio boxes with different dimensions.
- MPDIoU metric: MPDIoU directly minimizes distances between the top-left and bottom-right points of predicted and groundtruth rectangles.The metric is calculated from the coordinates of these corresponding corners.
- Loss formulation: LMPDIoU is defined as an MPDIoU-based bounding-box regression loss alongside the conventional IoU loss.During training, predicted boxes are optimized toward their groundtruth boxes by minimizing the selected loss.
- Geometric factors: MPDIoU incorporates nonoverlapping area, central-point distance, and width-and-height deviation using rectangle corner coordinates.This coordinate-based formulation simplifies the factors considered by existing loss functions.
- Same-aspect-ratio case: When boxes share an aspect ratio but differ in width and height, GIoU, DIoU, CIoU, and EIoU can assign equal values, whereas MPDIoU distinguishes them.For the stated construction, the central points overlap and MPDIoU(Bgt, Bprd1) > MPDIoU(Bgt, Bprd2).
- Loss properties: For arbitrary predicted coordinates, the construction ensures positive union area and bounds LMPDIoU by 0 ≤ LMPDIoU < 3.The groundtruth area is assumed positive, while predicted and intersection areas are constrained to be non-negative.
- Nonoverlapping case: For nonoverlapping boxes, minimizing LMPDIoU minimizes the normalized corner-point distance term.The loss is simplified when IoU = 0, and the cited analysis bounds the relevant distance expression.
4. Experimental Results
Experiments replace baseline bounding-box regression losses in YOLOv7 and YOLACT, evaluating detection, scene-text spotting, and instance segmentation across established benchmarks. The reported comparisons include accuracy, training behavior, and efficiency-related measures.
- Experimental setup: The study evaluates LMPDIoU against GIoU, DIoU, CIoU, and EIoU in YOLOv7, and against smooth L1 in YOLACT.The replacements are made in the models’ bounding-box regression components.
- Datasets: Experiments cover PASCAL VOC and MS COCO for detection, IIIT5K and MTHv2 for character-level scene text spotting, and MS COCO for instance segmentation.PASCAL VOC instance-segmentation results are omitted because the images are too small for strong performance.
- Evaluation: The reported object-detection measures include precision, mAP@0.5:0.95, and AP75.AP75 denotes mAP at an IoU threshold of 0.75.
- Object detection: Figure 5 compares YOLOv7 predictions on MS COCO 2017 and PASCAL VOC 2007 across five regression losses.The visualizations are ordered from LGIoU through LMPDIoU.
- Object detection: Figure 6 tracks YOLOv7 bounding-box loss and AP against training iterations on PASCAL VOC 2007&2012 for the same five losses.The figure compares training trajectories rather than only final test-set values.
- Instance segmentation: On YOLACT, LMPDIoU performs better than most existing losses on most reported metrics, while its improvement over competing losses is smaller than in earlier experiments.The authors associate the smaller gain with denser anchor boxes and limited remaining accuracy headroom, while noting efficiency room.
5. Conclusion
The paper introduces MPDIoU, a minimum-point-distance metric for bounding-box comparison, and its corresponding regression loss LMPDIoU. The authors report improved performance across object detection, scene-text spotting, and instance-segmentation benchmarks.
- Conclusion: MPDIoU compares arbitrary bounding boxes using minimum point distance while retaining appealing properties of existing IoU-based metrics and simplifying calculation.The paper presents LMPDIoU as the corresponding bounding-box regression loss.
- Conclusion: Applying LMPDIoU improves reported performance on PASCAL VOC, MS COCO, MTHv2, and IIIT5K using common measures and MPDIoU itself.The evaluated applications include object detection, scene-text spotting, and instance segmentation.
- Future work: The authors identify further downstream experiments, including scene text spotting and person re-identification, as future work.