[Paper Review] Alpha-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression

Jiabo He, Sarah Erfani+4 more|arXiv (Cornell University)|Oct 26, 2021

Advanced Neural Network ApplicationsComputer Science46 references144 citations

TL;DR

Alpha-IoU generalizes IoU-based losses into a power-based family with a single alpha parameter, improving bounding box regression and robustness across models/datasets.

ABSTRACT

Bounding box (bbox) regression is a fundamental task in computer vision. So far, the most commonly used loss functions for bbox regression are the Intersection over Union (IoU) loss and its variants. In this paper, we generalize existing IoU-based losses to a new family of power IoU losses that have a power IoU term and an additional power regularization term with a single power parameter $α$. We call this new family of losses the $α$-IoU losses and analyze properties such as order preservingness and loss/gradient reweighting. Experiments on multiple object detection benchmarks and models demonstrate that $α$-IoU losses, 1) can surpass existing IoU-based losses by a noticeable performance margin; 2) offer detectors more flexibility in achieving different levels of bbox regression accuracy by modulating $α$; and 3) are more robust to small datasets and noisy bboxes.

Motivation & Objective

• Generalize existing IoU-based bbox regression losses into a unified alpha-IoU family using a power transformation.
• Analyze key properties such as order preservation and loss/gradient reweighting under varying alpha.
• Demonstrate empirical improvements over standard IoU losses across multiple detectors and benchmarks.
• Show robustness to small datasets and noisy bounding boxes without adding training overhead.

Proposed method

• Define alpha-IoU loss via L_{alpha-IoU} = 1 - IoU^{alpha} (alpha > 0) and relate to existing losses (IoU, log(IoU)).
• Extend to a general form with an optional penalty term: L_{alpha-IoU} = 1 - IoU^{alpha} + P^{alpha}(B, B^{gt}).
• Unify existing IoU-based losses (IoU, GIoU, DIoU, CIoU) under the alpha-IoU framework by applying the same alpha to both IoU and penalty terms.
• Analyze properties: order preservingness, relative loss reweighting w_Lr = L_{alpha-IoU}/L_IoU, and relative gradient reweighting w_nabla_r.
• Provide empirical evaluation on PASCAL VOC and MS COCO with models like YOLOv5s/x and DETR, highlighting alpha=3 as a practical choice.
• Claim no additional model parameters or training time overhead are required.

Experimental results

Research questions

• RQ1 Does the alpha-IoU loss provide consistent improvements over standard IoU-based losses across different detectors and datasets?
• RQ2 How does the alpha parameter affect weighting of high-IoU objects in loss and gradient terms?
• RQ3 Is alpha-IoU robust to small datasets and noisy bounding boxes compared to traditional losses?
• RQ4 Can the alpha-IoU framework be applied to anchor-based and anchor-free detectors without architectural changes?

Key findings

• Alpha-IoU with alpha > 1 upweights high-IoU objects, improving bbox regression accuracy and high-precision mAP (e.g., mAP_{75:95}).
• Empirical gains observed across YOLOv5s, YOLOv5x, and DETR on PASCAL VOC and MS COCO, especially at higher IoU thresholds (AP_{95}).
• Alpha=3 provides strong and consistent performance gains across models/datasets, with limited sensitivity to dataset/model choice.
• The method improves robustness on small training sets and noisy bboxes without changing training procedures or adding parameters.