search

Walter0807 / MotionBERT

[ICCV 2023] PyTorch Implementation of "MotionBERT: A Unified Perspective on Learning Human Motion Representations"
Apache License 2.0
1k stars 120 forks source link

simple question on action recognition fine tuning data prep #106

Closed valentin-fngr closed 9 months ago

valentin-fngr commented 9 months ago

Hi,

I am currently building a dataset in order to fine tune the model for my own action rec taskl. I have extracted keypoints using Alphapose as mentioned in inference_wild.py using :

1) Person detection 2) HRNET

Looking at the NTU60 dataset, 1020 x 1080 seem to be there input image shape. But, my inputs will probably be smaller due to the bounding box extraction.

Would you advise to train using the cropped person as input or the entire image containing the image ?

Best

valentin-fngr commented 9 months ago

Closing as it seems like I should not used the cropped image as Input but the entire frame