search

lijiaman / egoego_release

Official Implementation of the Paper: Ego-Body Pose Estimation via Ego-Head Pose Estimation (CVPR 2023)
MIT License
120 stars 10 forks source link

Running test_egoego_pipeline.sh over EPIC Kitchens #2

Closed edessa closed 9 months ago

edessa commented 1 year ago

Hello,

I am trying to extract full body pose information from the EPIC Kitchens dataset by running the script you all have provided. I must beforehand extract head pose data from DROID-SLAM and optical flow from RAFT. I have had trouble producing the exact optical flow features your pipeline requires, as the RAFT repository does not provide models that produce the optical flow dimensions (512,) your pipeline requires. Could you please point to which of the models on the RAFT repository you ended up using, and which variable within the model architectures you use to retrieve the features? Thank you ahead of time!

L-Scofano commented 1 year ago

Hello, I'm currently facing the same problem with RAFT, which model should be used? Thank you in advance!

lijiaman commented 9 months ago

Hi, We used ResNet to extract the optical flow features. You can find the ResNet extractor here https://github.com/lijiaman/egoego_release/blob/main/egoego/model/resnet.py#L25. For using RAFT, I believe we used a default one, either network should work.