mit-han-lab / temporal-shift-module

[ICCV 2019] TSM: Temporal Shift Module for Efficient Video Understanding
https://arxiv.org/abs/1811.08383
MIT License
2.07k stars 417 forks source link

TSN only uses 1 segment? #144

Closed zhengshou closed 4 years ago

zhengshou commented 4 years ago

Hi, maybe I am missing something, but according to log like the below one, looks like when using TSN 2D model, only 1 test_segment, i.e. 1 frame is used? I thought it is supposed to average results from input frames, say 8 frames?

https://file.lzhu.me/projects/tsm/models/log/testlog_uniform_TSM_kinetics_RGB_resnet50_avg_segment5_e50.log

Thanks!

tonylins commented 4 years ago

Hi, sorry for the confusion. The num_segments parameter is used only when you enable TSM or Non-local in this codebase. For TSN, the parameter is not considered. But during testing, we feed 8 frames into the TSN model.

zhengshou commented 4 years ago

Got it, thanks Ji!