Boosting person ReID feature extraction via dynamic convolution

Abstract

AbstractExtraction of discriminative features is crucial in person re-identification (ReID) which aims to match a query image of a person to her/his images, captured by different cameras. The conventional deep feature extraction methods on ReID employ CNNs with static convolutional kernels, where the kernel parameters are optimized during the training and remain constant in the inference. This approach limits the network's ability to model complex contents and decreases performance, particularly when dealing with occlusions or pose changes. In this work, to improve the performance without a significant increase in parameter size, we present a novel approach by utilizing a channel fusion-based dynamic convolution backbone network, which enables the kernels to change adaptively based on the input image, within two existing ReID network architectures. We replace the backbone network of two ReID methods to investigate the effect of dynamic convolution on both simple and complex networks. The first one called Baseline, is a simpler network with fewer layers, while the second, CaceNet represents a more complex architecture with higher performance. Evaluation results demonstrate that both of the designed dynamic networks improve identification accuracy compared to the static counterparts. A significant increase in accuracy is reported under occlusion tested on Occluded-DukeMTMC. Moreover, our approach achieves a performance comparable to the state-of-the-art on Market1501, DukeMTMC-reID, and CUHK03 with a limited computational load. These findings validate the effectiveness of the dynamic convolution in enhancing the person ReID networks and push the boundaries of performance in this domain.