search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
419 stars 77 forks source link

ActivityNet 1-class vs. 200-class problem #40

Closed egoodman92 closed 2 years ago

egoodman92 commented 2 years ago

Hello - thanks for the great contribution. Was wondering, for ActivityNet1.3, is the typical benchmark for the one-class problem? Saw that the config for activitynet config there was "num_classes: 1", and trying to understand if that is standard versus the 200-class problem. (can't seem to find confirmation in other literature) or I am misunderstanding the codebase.

Thanks!!

tzzcl commented 2 years ago

Thank you for your interest in our project.

You are right. In ActionFormer, we directly predict proposals for ActivityNet v1.3 and fuse the proposal with the external classification scores. If we directly learn with the 200-class problem, we will face an extreme imbalance learning problem. Many existing works also tries to learn proposals for ActivityNet v1.3 like G-TAD

tzzcl commented 2 years ago

Closed due to inactivity.