search

HumamAlwassel / TSP

TSP: Temporally-Sensitive Pretraining of Video Encoders for Localization Tasks (ICCVW 2021)
http://humamalwassel.com/publication/tsp/
MIT License
107 stars 16 forks source link

Transfer to G-TAD with TSP extracted features. #12

Closed bofang98 closed 2 years ago

bofang98 commented 2 years ago

Hi, @HumamAlwassel Thanks for your excellent work and for sharing the code. I'm new in temporal action detection and temporal proposal generation areas. Now I want to test my pretraining 3DResNet18 models' generation ability on the above two tasks. After extracting thumos features, how to validate its effectiveness when transferred to G-TAD, BMN framework (In other words, how to handle the features further)? If it is convenient for you, would you mind sharing your change on the source code of G-TAD or BMN? My email is fangbo.official@foxmail.com Looking forward to your reply.

HumamAlwassel commented 2 years ago

Hi @BoFang13,

Thanks for your interest in TSP and welcome to the TAL community :) We do not change anything in the code of the algorithms we use. We simply swap the features they use with our TSP features. However, we post-process our features (by linear interpolation) to be of the same number of features per video as the original features used in G-TAD, BMN, ... etc. We do this to avoid changing anything in the code of these algorithms (e.g. G-TAD uses hard-coded frame-level annotations based on the original features they use).

Please take a look at these two issues where I explain the post-processing in details: issue 8 and issue 9.

Cheers!