Revisiting the Loss Weight Adjustment in Object Detection

TL;DR

This paper introduces ALWA, an adaptive method for balancing classification and regression losses in object detection, leading to improved performance across datasets by addressing task imbalance.

Contribution

We propose ALWA, a novel adaptive loss weight adjustment method that enhances multi-task learning in object detection by considering statistical loss characteristics.

Findings

Significant performance improvements on PASCAL VOC and MS COCO datasets.

Effective with various regression loss functions like L1, SmoothL1, and CIoU.

Addresses the imbalance between classification and regression tasks in anchor-based detectors.

Abstract

Object detection is a typical multi-task learning application, which optimizes classification and regression simultaneously. However, classification loss always dominates the multi-task loss in anchor-based methods, hampering the consistent and balanced optimization of the tasks. In this paper, we find that shifting the bounding boxes can change the division of positive and negative samples in classification, meaning classification depends on regression. Moreover, we summarize three important conclusions about fine-tuning loss weights, considering different datasets, optimizers and regression loss functions. Based on the above conclusions, we propose Adaptive Loss Weight Adjustment(ALWA) to solve the imbalance in optimizing anchor-based methods according to statistical characteristics of losses. By incorporating ALWA into previous state-of-the-art detectors, we achieve a significant performance gain on PASCAL VOC and MS COCO, even with L1, SmoothL1 and CIoU loss. The code is available at https://github.com/ywx-hub/ALWA.