Loss reweight in scale dimension: A simple while effective feature selection strategy for anchor-free detectors

Abstract

Feature selection plays an important role during training in object detection and it is becoming a hotspot these years. Here we present a simple yet effective feature selection strategy called Loss Reweight in Scale Dimension(LRSD) and train a single-stage anchor-free detector, termed LRD. To optimize the heuristic-guided feature selection process across FPN levels, for each instance, LRD dynamically reweights the training loss of positive samples from selected top-k feature levels by introducing a reweight function. Because of its rigorous and nonlinear mathematical properties, a precise sample loss reweighting procedure across scale dimension could be done. Without adding extra meta-nets or branches, LRD improves detection performance economically without sacrificing inference speed. Moreover, our detection framework can be further improved by recently proposed transformer-based feature extraction networks such as swin-transformer. Extensive experiments show that LRD achieves 40.4% AP at a speed of 16.7 fps with ResNet-50 as the backbone and helps to improve detection performance by around 1.9% ~ 2.6% compared with our baseline, symbolic one-stage anchor-free detector Foveabox using ResNet-101 as the backbone. Codes are released at ( https://github.com/PanffeeReal/LRSD-LRD ). Fig 2. Training and inference pipeline of LRD.