Multi-stage Adaptive Rank Statistic Pruning for Lightweight Human 3D Mesh Recovery Model

Abstract

We present a rank statistic adaptive multi-stage pruning method to find lightweight neural networks for 3D human mesh recovery while minimizing accuracy drop. We observe that some feature maps often have prominent low-rank patterns regardless of input human images. Furthermore, even after pruning, feature channels that should have been pruned according to pruning criteria frequently re-appear in test time. From these observations, we design rank statistic adaptive multi-stage pruning; thereby, we can prune more filters with recovering mesh reconstruction accuracy. We demonstrate that, for DenseNet-121, 60.0% of parameters and 67.9% of FLOPs are saved while maintaining comparable accuracy to that of the original full model. This is a notable improvement compared to the competing method based on the L1 filter pruning, where the error is increased by 17.55% at the same pruning rate.