Custom I3D models

Forked from the original repo of "Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset" paper by Joao Carreira and Andrew Zisserman.

Take a look at the original README.md here.

Custom files

This repo aims at widening the original use of Deepmind's I3D repo which gives only logits and probabilities on sample videos. Furthermore, it tests a continuous run of the I3D on entire datasets (note that this can be greatly improved).

Example use

python custom.py --imagenet_pretrained true --eval_type joint --final_endpoint Mixed_4f --path path --flow_path path --save_path path --frames 99