thomasverelst / dynconv

Code for Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference (CVPR2020)
https://arxiv.org/abs/1912.03203
126 stars 14 forks source link
conditional-computation conditional-execution deep-learning dynamic-convolutions dynamic-execution neural-networks pytorch

Dynamic Convolutions - DynConv

Pytorch code for DynConv. DynConv applies convolutions on important regions of the image only, and thus reduces the computational cost while speeding up inference up to 2 times.

https://arxiv.org/abs/1912.03203

Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference
Thomas Verelst and Tinne Tuytelaars
CVPR 2020

Note that our work consists of two parts:

The first point is demonstrated on both classification and human pose estimation, the second point only on human pose estimation.

Classification

Human Pose Estimation

Coming later

Teaser GIF

Click thumbnail for 1-minute Youtube overview video