Source-linked AI summary
Focaler-IoU: More Focused Intersection over Union Loss
Hao Zhang, Shuaijie Zhang
TL;DR
Object-detection localization depends heavily on bounding-box regression losses, yet prior geometric approaches do not account for difficult and simple sample distributions. The paper analyzes this distribution, proposes Focaler-IoU using linear interval mapping, and reports improved detector performance across different detection tasks.
Problem
Existing bounding-box regression research uses box geometry while ignoring the impact of difficult and simple sample distributions on regression results.
Method
Focaler-IoU reconstructs the original IoU loss through linear interval mapping to focus regression on different sample types.
Results
Comparative experiments with advanced detectors and regression methods report improved detection performance from Focaler-IoU across different detection tasks.
Takeaways & Limitations
Focaler-IoU provides task-dependent focus on difficult or simple regression samples, according to their distribution.
Abstract
from arXiv · showhide
Bounding box regression plays a crucial role in the field of object detection, and the positioning accuracy of object detection largely depends on the loss function of bounding box regression. Existing researchs improve regression performance by utilizing the geometric relationship between bounding boxes, while ignoring the impact of difficult and easy sample distribution on bounding box regression. In this article, we analyzed the impact of difficult and easy sample distribution on regression results, and then proposed Focaler-IoU, which can improve detector performance in different detection tasks by focusing on different regression samples. Finally, comparative experiments were conducted using existing advanced detectors and regression methods for different detection tasks, and the detection performance was further improved by using the method proposed in this paper.Code is available at \url{https://github.com/malagoutou/Focaler-IoU}.
I. INTRODUCTION
Bounding box regression loss is central to object localization, but existing approaches largely emphasize box geometry rather than difficult-versus-simple sample distributions. The paper introduces Focaler-IoU to focus regression on different sample types and reports improved detection performance.
- Bounding box regression loss is an important component of the localization branch in anchor-based and anchor-free detectors.
- IoU-based losses improve detection accuracy but have shortcomings, including vanishing gradients when ground-truth and anchor boxes do not overlap.
- Existing losses incorporate geometric constraints such as enclosing-box, centroid-distance, and aspect-ratio relationships between boxes.
- Training samples can be difficult or simple, and the useful focus depends on their distribution within the detection task.
- Focaler-IoU reconstructs the original IoU loss through linear interval mapping to focus on different regression samples.
- Comparative experiments with advanced one-stage detectors verify that the method can improve detection performance and compensate for shortcomings of existing methods.
II. RELATED WORK
The related work traces bounding-box regression losses from IoU through extensions that add geometric constraints. IoU is presented as a widely used detection criterion, while later methods address its limitations.
- IoU and successive losses including GIoU, DIoU, CIoU, EIoU, and SIoU form a progression of bounding-box regression methods.
- IoU is described as the most popular target-detection evaluation criterion.
- The IoU metric is defined from the intersection and union of a predicted box and a ground-truth box.
B. GIoU Metric
GIoU, DIoU, and CIoU extend IoU by adding geometric information for non-overlap handling, centroid distance, and shape similarity. These constraints target known localization and convergence limitations.
- B. GIoU Metric: GIoU addresses IoU’s no-overlap gradient-vanishing problem using the smallest enclosing box of the ground-truth and anchor boxes.
- C. DIoU Metric: DIoU adds a centroid normalized distance loss term to IoU, with distance measured between the anchor and ground-truth box centers.
- C. DIoU Metric: CIoU extends DIoU with a shape loss term that reduces aspect-ratio differences between anchor and ground-truth boxes.
D. EIoU Metric
EIoU redefines the shape loss used with CIoU and directly reduces aspect differences between ground-truth and anchor boxes. Its enclosing-box dimensions provide the width and height terms for that comparison.
- D. EIoU Metric: EIoU redefines CIoU’s shape loss and directly reduces aspect differences between ground-truth and anchor boxes.
- D. EIoU Metric: The minimum box covering the ground-truth and anchor boxes supplies the width and height denoted by wc and hc.
E. SIoU Metric
SIoU incorporates the angle between bounding boxes to accelerate convergence by reducing the angle toward horizontal or vertical alignment. Figure 1 maps difficult and simple samples with separate linear interval curves.
- E. SIoU Metric: SIoU considers the angle between anchor and ground-truth boxes to accelerate convergence.Its stated aim is to decrease the angle toward the horizontal or vertical direction.
- E. SIoU Metric: Figure 1 presents linear interval mapping curves for difficult and simple samples in panels (a) and (b), respectively.
A. Analysis
Object detection regression involves imbalanced difficult and simple samples, and the useful focus depends on which sample type dominates the task. Simple targets are generally easier, whereas extremely small targets are difficult to position precisely.
- A. Analysis: Imbalanced detection samples can be categorized as difficult or simple according to detection difficulty.
- A. Analysis: General-scale targets are treated as simple samples, while extremely small targets are difficult because precise positioning is challenging.
- A. Analysis: Focusing bounding box regression on simple samples may improve performance when simple samples dominate the detection task.
- A. Analysis: The passage indicates that difficult-sample-dominated tasks require a different regression focus than tasks with many simple samples.
B. Focaler-IoU
Focaler-IoU reconstructs IoU loss through linear interval mapping so regression can focus on different sample types and improve edge regression. The reconstructed loss is then applied to existing IoU-based bounding box losses.
- B. Focaler-IoU: Focaler-IoU uses linear interval mapping to reconstruct IoU loss and focus on different regression samples.The stated purpose is to support different detection tasks and improve edge regression.
- B. Focaler-IoU: The reconstruction uses bounds [d, u] within [0, 1], with their values controlling which regression samples receive focus.IoU focaler denotes the reconstructed Focaler-IoU, while IoU is the original value.
- B. Focaler-IoU: Focaler-IoU is applied to existing IoU-based losses, producing Focaler-GIoU, Focaler-DIoU, Focaler-CIoU, Focaler-EIoU, and Focaler-SIoU.
IV. EXPERIMENTS
The experiments use PASCAL VOC with YOLOv8s and YOLOv7-tiny, comparing SIoU against the proposed Focaler-SIoU. Table I reports their YOLOv8 performance.
- IV. EXPERIMENTS: The PASCAL VOC experiment trains on 16,551 images from VOC2007 and VOC2012 and tests on 4,952 VOC2007 images.
- IV. EXPERIMENTS: YOLOv8s and YOLOv7-tiny are used as one-stage detectors for comparison on the VOC dataset.
- IV. EXPERIMENTS: The experiment compares SIoU with the proposed Focaler-SIoU regression method.
- IV. EXPERIMENTS: Table I presents the performance of SIoU and Focaler-SIoU on YOLOv8.
B. AI-TOD on YOLOv5
On AI-TOD, a tiny-target remote-sensing dataset, YOLOv5s was evaluated with SIoU and Focaler-SIoU; the paper reports improved detection performance from its proposed approach.
- B. AI-TOD on YOLOv5: AI-TOD contains many tiny targets averaging 12.8 pixels, making it a difficult localization setting.The experiment uses YOLOv5s with SIoU as the comparison method.
- B. AI-TOD on YOLOv5: The experiment compares SIoU and Focaler-SIoU on YOLOv5.The corresponding results are reported in Table II.
- B. AI-TOD on YOLOv5: Focaler-IoU reconstructs the original IoU loss through linear interval mapping to focus on difficult or easy samples.The targeted sample type depends on whether difficult or simple samples dominate.
- B. AI-TOD on YOLOv5: Comparative experiments show that the proposed method can effectively improve detection performance.