Bag of Freebies for Training Object Detection Neural Networks
Zhi Zhang, Tong He, Hang Zhang, Zhongyue Zhang, Junyuan Xie, Mu Li Amazon Web Services {zhiz, htong, hzaws, zhongyue, junyuanx,
Abstract
Training heuristics greatly improve various image classification model accuracies [8]. Object detection models, however, have more complex neural network structures and optimization targets. The training strategies and pipelines dramatically vary among different models. In this works, we explore training tweaks that apply to various models including Faster R-CNN and YOLOv3. These tweaks do not change the model architectures, therefore, the inference costs remain the same. Our empirical results demonstrate that, however, these freebies can improve up to 5% absolute precision compared to state-of-the-art baselines.
中文速览
目标检测模型种类繁多、训练流程各异，如何在不改动网络结构的前提下统一提升各类检测器的精度，是一个被长期忽视的问题。研究者系统梳理并验证了一批"免费午餐"式训练技巧，核心包括：专为检测任务设计的视觉连贯图像混合增强（Visually Coherent Mixup）、分类头标签平滑、余弦学习率调度搭配学习率预热，以及针对单阶段与多阶段检测器差异化定制的数据预处理策略。将这些技巧叠加应用于 Faster R-CNN 和 YOLOv3 后，在不增加任何推理开销的情况下，模型精度相比强基线最高提升了 5 个百分点（相对提升约 15%–20%）。这项工作首次系统地将图像分类领域的训练经验迁移并适配到目标检测场景，为后续研究者提供了一套可直接复用的实践指南。
原文 arXiv:1902.04103;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1902.04103v3