基于渐进式训练的多判别器域适应目标检测

Domain Adaptive Multi-Discriminator Object Detection Based on Progressive Training

基于对抗训练的域适应目标检测的研究旨在不对新数据集进行额外标注的情况下,将检测模型应用于不同的数据集。但现有算法存在目标检测和域对齐任务难以平衡的问题,且一般的单判别器结构容易局限于数据的单个模式,导致域对齐的质量下降。提出一种基于渐进式训练的多判别器域适应目标检测算法,针对传统的单判别器结构对复杂结构数据进行域对齐时的局限性,在实例级的域适应头中引入多判别器结构,使其在学习域不变信息时考虑数据的多模结构,实现质量更高、更全面的域对齐。同时,为降低引入多判别器结构而增加的模型复杂度,设计基于Dropout技术的多判别器结构,对单个判别器参数进行重复利用,并创新性地引入渐进式训练策略,即随着训练的推进逐步增大域对齐任务的比重和难度,动态平衡目标检测和域对齐任务的权重。实验结果表明,所提算法在Cityscapes到Foggy Cityscapes的域适应场景下的平均检测精度为42.9%,相比近几年该领域的新算法提高了至少0.5个百分点。

The research on domain adaptive object detection based on adversarial training aims the deployment of detection models for use with different data sets without labeling new data sets.However,existing algorithms have difficulty in balancing the tasks of object detection and domain alignment.The general single discriminator structure is limited to single mode of data,resulting in degradation of domain alignment quality.This paper proposes a multi-discriminator domain adaptive object detection algorithm based on progressive training.Considering the limitations of the traditional single-discriminator structure in the domain alignment of complex structural data,a multi-discriminator structure is introduced into the instance-level domain-adaptive head to force it to consider multiple modes of data while learning the domain invariant information,which contributes to achieving higher quality and more comprehensive domain alignment.Meanwhile,to reduce the increased model complexity,a multi-discriminator structure designed to reuse the single discriminator parameters is introduced based on dropout technology.In this paper,an innovative progressive training strategy is introduced,whereby the proportion and difficulty of domain alignment are gradually increased with the progress in training,to dynamically balance the weight of object detection and domain alignment tasks.The experimental results indicate that the mean average precision of the algorithm in domain adaptation from Cityscapes to Foggy Cityscapes was 42.9%,which is an improvement of at least 0.5 percentage points compared to algorithms of recent years.