search

Walter0807 / MotionBERT

[ICCV 2023] PyTorch Implementation of "MotionBERT: A Unified Perspective on Learning Human Motion Representations"
Apache License 2.0
1.02k stars 123 forks source link

Embeddings #78

Closed moss-co closed 1 year ago

moss-co commented 1 year ago

Hi, I have been trying to use train_action_1shot.py to get those using this model config, but challenged by the machine config.

Is there any any other approach you recommend we get the embeddings ? I also see that the checkpoints for mainly the model params. How can I get the embeddings from these checkpoints ?

Also, perhaps do you have the pre-trained embeddings available?

Walter0807 commented 1 year ago

Hi, do you want to obtain the pretrained motion representations or do you want to fine-tune MB for one-shot action recognition?

Walter0807 commented 1 year ago

hey - was Iooking for pre-trained motion representations.

https://github.com/Walter0807/MotionBERT#using-motionbert-for-human-centric-video-representations This section should be what you are looking for.

Walter0807 commented 1 year ago

Also in ActionNet, input dim ( Input: (N, M, T, J, C)) what does M represent ?

It's the dimension of persons (M=num of persons).

moss-co commented 1 year ago

thank you for clarifying! appreciate it. Should we expect the network to learn similar embeddings/feature representations when training both ActionNet and MeshRegressor from just the backbone without the head ?

I think I am also not clear on what xsubject and xview mean here (giving M = 2) , I assumed we are working with single human per input pose representations ?