Journal
MULTIMEDIA TOOLS AND APPLICATIONS
Volume -, Issue -, Pages -Publisher
SPRINGER
DOI: 10.1007/s11042-023-16731-w
Keywords
Person re-identification; Filter pruning; Optimization problem; Evolutionary algorithm
Ask authors/readers for more resources
In this paper, a filter pruning scheme based on evolutionary algorithms (EAFPruner) is proposed for person re-identification on resource-constrained platforms. The optimal pruning structure in the pruning space is automatically found by balancing computational cost and identification accuracy. Experimental results show the effectiveness of the proposed method by reducing the number of parameters while maintaining comparable accuracy.
Filter pruning has drawn much attention for releasing the limitations of person re-identification (ReID) in resource-constrained platforms due to the large parameters and high computational consumption. Filter pruning is depended on the given combination of pruning ratios to trim off unimportant filters, therefore, we define filter pruning as optimization problem. Existing strategies generally retrain all the pruning results and compare their performance iteratively, representing a major expenditure of time and effort. In this paper, we formally put forward a scheme of filter pruning based on evolutionary algorithms (EAFPruner) for ReID task for the first time, which automatically finds the optimal pruning structure in the pruning space with the better trade-off between computational cost and identification accuracy. Specifically, multiple evolutionary algorithms are adopted to solve the optimization problem by using mean average precision (mAP) as the fitness for filter-level sparsity. In addition, the adaptive batch normalization strategy is adopted to evaluate the pruned candidates generated by evolutionary algorithm without fine-tuning them to speed up training process. The experimental results are carried out on four person ReID datasets, and the number of parameters required by ResNet50 architecture is reduced by half while maintaining a comparable Rank-1 accuracy, which demonstrates the effectiveness of our proposed method.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available