A Flow-based Attention Lightweight Network for Human Pose Estimation. Our network FALNet-50 only has 2.2M parameters and 0.66G FLOPs, achieving 70.4 in AP score on COCO val2017 with 16 FPS inference speed on a smartphone. The code is about to be open source.
12-10-8 / falnet Goto Github PK
View Code? Open in Web Editor NEWA Flow-based Attention Lightweight Network for Human Pose Estimation
License: Apache License 2.0