[论文解读] A Survey of Pruning Methods for Efficient Person Re-identification Across Domains

Hugo Masson, Amran Bhuiyan+5 more|arXiv (Cornell University)|Jul 4, 2019

Video Surveillance and Tracking Methods参考文献 22被引用 9

一句话总结

本文综述了用于压缩跨域行人重识别中深度孪生网络的剪枝技术，表明结构化剪枝可将ResNet特征提取器的FLOPS减少50%，同时将排名-1准确率保持在原始模型的1%以内。该方法可在资源受限平台实现高效部署，且准确率损失可忽略不计。

ABSTRACT

Recent years have witnessed a substantial increase in the deep learning architectures proposed for visual recognition tasks like person re-identification, where individuals must be recognized over multiple distributed cameras. Although deep Siamese networks have greatly improved the state-of-the-art accuracy, the computational complexity of the CNNs used for feature extraction remains an issue, hindering their deployment on platforms with with limited resources, or in applications with real-time constraints. Thus, there is an obvious advantage to compressing these architectures without significantly decreasing their accuracy. This paper provides a survey of state-of-the-art pruning techniques that are suitable for compressing deep Siamese networks applied to person re-identification. These techniques are analysed according to their pruning criteria and strategy, and according to different design scenarios for exploiting pruning methods to fine-tuning networks for target applications. Experimental results obtained using Siamese networks with ResNet feature extractors, and multiple benchmarks re-identification datasets, indicate that pruning can considerably reduce network complexity while maintaining a high level of accuracy. In scenarios where pruning is performed with large pre-training or fine-tuning datasets, the number of FLOPS required by the ResNet feature extractor is reduced by half, while maintaining a comparable rank-1 accuracy (within 1\% of the original model). Pruning while training a larger CNNs can also provide a significantly better performance than fine-tuning smaller ones.

研究动机与目标

• 分析适用于压缩行人重识别中深度孪生网络的最先进剪枝技术。
• 根据剪枝标准和目标应用场景的设计场景，评估剪枝策略。
• 评估剪枝对跨域行人重识别中计算复杂度和准确率的影响。
• 比较在较小模型上训练期间剪枝与微调期间剪枝的性能表现。

提出的方法

• 本文综述了专为基于ResNet的特征提取器的孪生网络量身定制的各种剪枝技术。
• 基于重要性评分标准（如权重大小、梯度和激活值）评估剪枝方法。
• 通过结构化剪枝减少ResNet主干网络的FLOPS，同时保持特征表示质量。
• 在多个行人重识别基准数据集上进行实验，以评估准确率与效率之间的权衡。
• 在训练过程中以及在预训练模型上进行微调时应用剪枝，以比较性能结果。
• 评估大规模预训练数据集和微调数据集对剪枝效果的影响。

实验结果

研究问题

• RQ1 在基于ResNet的孪生网络中，剪枝在减少FLOPS方面的有效性如何？
• RQ2 在大型预训练模型上应用剪枝时，准确率下降程度如何？
• RQ3 与微调较小模型相比，训练期间剪枝的性能表现如何？
• RQ4 哪种剪枝标准在不同数据集上能实现最稳定的准确率保留？
• RQ5 剪枝是否能在不造成显著准确率损失的前提下，实现资源受限平台上的实时部署？

主要发现

• 剪枝使ResNet特征提取器的FLOPS减少50%，同时将排名-1准确率保持在原始模型的1%以内。
• 在大规模预训练或微调数据集的场景下，剪枝可实现显著的FLOPS减少，且准确率下降不明显。
• 训练期间剪枝的性能优于微调较小模型，表明训练阶段剪枝更具有效性。
• 准确率在多个行人重识别基准上保持稳定，证明了该剪枝方法的鲁棒性。
• 结果证实，结构化剪枝可实现边缘设备上的高效部署，满足实时性要求。
• 该方法即使在大幅压缩后仍能保持判别性特征学习能力，验证了其在实际部署中的适用性。