How to implement on a custom video?

abduallahmohamed / Social-STGCNN

Code for "Social-STGCNN: A Social Spatio-Temporal Graph Convolutional Neural Network for Human Trajectory Prediction" CVPR 2020

MIT License

489 stars 140 forks source link

Closed Mzunoven closed 4 years ago

Mzunoven commented 4 years ago

Thanks for your great work.

I want to ask how to implement this network on a custom video? Can you give some instructions on how to prepare my video and do prediction?

Thanks!

abduallahmohamed commented 4 years ago

I originally didn't process the data as I used ETH & UCY datasets as it's. If I were in your place I would:

Sample the video at fixed periods (0.2 seconds for example).
Annotate the pedestrians in the video while having a fixed reference.
Construct the homography matrix that converts from pixel domain to meters(real-world) domain.
Generate the annotation same way as eth & ucy files are.