[Paper Review] Focal and Efficient IOU Loss for Accurate Bounding Box Regression

Yifan Zhang, Weiqiang Ren+4 more|arXiv (Cornell University)|Jan 20, 2021

Advanced Neural Network ApplicationsComputer Science65 citations

TL;DR

The paper introduces Efficient IOU (EIOU) loss and a regression version of focal loss, then combines them as Focal-EIOU to accelerate convergence and improve bounding box localization in object detection.

ABSTRACT

In object detection, bounding box regression (BBR) is a crucial step that determines the object localization performance. However, we find that most previous loss functions for BBR have two main drawbacks: (i) Both $\ell_n$-norm and IOU-based loss functions are inefficient to depict the objective of BBR, which leads to slow convergence and inaccurate regression results. (ii) Most of the loss functions ignore the imbalance problem in BBR that the large number of anchor boxes which have small overlaps with the target boxes contribute most to the optimization of BBR. To mitigate the adverse effects caused thereby, we perform thorough studies to exploit the potential of BBR losses in this paper. Firstly, an Efficient Intersection over Union (EIOU) loss is proposed, which explicitly measures the discrepancies of three geometric factors in BBR, i.e., the overlap area, the central point and the side length. After that, we state the Effective Example Mining (EEM) problem and propose a regression version of focal loss to make the regression process focus on high-quality anchor boxes. Finally, the above two parts are combined to obtain a new loss function, namely Focal-EIOU loss. Extensive experiments on both synthetic and real datasets are performed. Notable superiorities on both the convergence speed and the localization accuracy can be achieved over other BBR losses.

Motivation & Objective

• Address slow convergence and inaccuracy of existing BBR losses by improving geometric discrepancy measurement.
• Mitigate anchor box imbalance by designing a regression version of focal loss to emphasize high-quality anchors.
• Develop and validate a combined Focal-EIOU loss across synthetic and real datasets.
• Demonstrate improvements across state-of-the-art detectors on COCO 2017 dataset.

Proposed method

• Propose EIOU loss that decomposes L_EIOU into IOU loss plus distance and aspect components, explicitly minimizing overlap, center distance, and width/height differences.
• Design FocalL1 loss to reweight regression gradients, controlled by parameters beta and alpha to emphasize high-quality examples.
• Combine EIOU with focal weighting into Focal-EIOU loss using IOU-based reweighting to focus on informative anchors (L_Focal-EIOU = IOU^gamma * L_EIOU).
• Normalize batch weights to stabilize training (L_Focal-EIOU = sum(W_i * L_EIOU_i)/sum(W_i)).
• Evaluate on synthetic setups and COCO 2017 across multiple backbones/detectors (Faster R-CNN, Mask R-CNN, RetinaNet, ATSS, PAA, DETR).
• Perform ablations to isolate effects of EIOU, FocalL1, and reweighting strategies.

Experimental results

Research questions

• RQ1 Does EIOU improve convergence speed and localization accuracy over IOU-based losses?
• RQ2 Can a regression-focused focal mechanism (EEM) adequately balance high- vs low-quality anchor contributions in BBR?
• RQ3 Does integrating FocalL1 with EIOU (Focal-EIOU) yield consistent gains across diverse detectors and backbones on COCO?
• RQ4 What are the effects of hyperparameters (beta, gamma) on performance and training stability?

Key findings

• EIOU achieves faster convergence and improved localization compared with IOU, GIOU, and CIOU losses in simulations and COCO experiments.
• Focal-L1-based reweighting boosts gradients from higher-quality anchors, enhancing AP by notable margins in ablations.
• Focal-EIOU consistently improves AP across various detectors on COCO 2017, with the best single-row result of 37.5 AP (56.1 AP50, 40.0 AP75, 21.1 AP_S, 40.9 AP_M, 49.8 AP_L).
• Table-driven ablations show Focal-EIOU outperforms baseline and other IOU-based losses, with gains of up to ~1.6 AP over baseline.
• Focal-EIOU improves localization especially for medium and large objects, while maintaining stability under appropriate gamma (gamma=0.5) and beta settings.