面向行人重识别的多域批归一化问题研究

Research on Batch Normalization for Multi-domain Person Re-identification

近年来基于深度神经网络的行人重识别算法取得了长足的进步,被广泛应用于网络中的批归一化(batch normalization)模块发挥着重要作用。批归一化模块在多数情况下可有效提高网络收敛速度和训练稳定性,然而当多个独立标注的数据库混合在一块进行跨域或者多域训练时,数据之间的分布差异使得目前的批归一化算法工作逻辑存疑。由于不同批次下训练数据的分布差异较大,归一化过程中的统计参数不稳定导致批归一化效果恶化。该文聚焦于多数据集合并下的行人重识别模型训练问题,通过对多数据集分布差异导致的多域模型批归一化存在的问题进行分析。然后针对模型批量归一化算法面对的多域差异,提出了一种解决策略,在多个数据集并行训练下提高了模型的泛化能力。实验结果表明:所提出的多域归一化方法在多域训练下能有效提高模型最终的泛化能力,获得更高的识别准确度,并且可应用于其他行人重识别网络以进一步提升模型性能。

In recent years,the pedestrian re-identification(Re-ID)algorithms based on deep neural networks(DNNs)have made considerable progress.The use of batch normalization modules in constructing DNNs is key to the success of deep Re-ID,which can effectively improve the convergence speed and training stability in most cases.However,when multiple independently labeled datasets are mixed together for cross-domain or multi-domain training,the domain gap makes the use of the current batch normalization module questionable.The distribution difference among training batches may result in unstable estimates for various statistical parameters,which may lead to possible performance deterioration for the employed batch normalization.We focus on the training of pedestrian re-recognition models under the combination of multiple data sets and analyze the problems in the batch normalization of multi-domain models caused by the distribution differences of multiple data sets.Aiming at the multi-domain difference of the batch model normalization algorithm,a strategy is proposed to improve the generalization ability of the model under the parallel training of multiple data sets.Experiment shows that the proposed domain-specific batch normalization algorithm can effectively improve the final generalization ability under multi-domain training and obtain higher recognition accuracy.Furthermore,it can be applied to existing Re-ID networks for further improving their performance.