This repository contains trained models reported in the paper "Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset" by Joao Carreira and Andrew Zisserman.
This code is based on Deepmind's Kinetics-I3D. Including PyTorch versions of their models.
We provide code to extract I3D features and fine-tune I3D for charades. Our fine-tuned models on charades are also available in the models director (in addition to Deepmind's trained models). The deepmind pre-trained models were converted to PyTorch and give identical results (flow_imagenet.pt and rgb_imagenet.pt). These models were pretrained on imagenet and kinetics (see Kinetics-I3D for details).
train_i3d.py contains the code to fine-tune I3D based on the details in the paper and obtained from the authors. Specifically, this version follows the settings to fine-tune on the Charades dataset based on the author's implementation that won the Charades 2017 challenge. Our fine-tuned RGB and Flow I3D models are available in the model directory (rgb_charades.pt and flow_charades.pt).
This relied on having the optical flow and RGB frames extracted and saved as images on dist. charades_dataset.py contains our code to load video segments for training.
extract_features.py contains the code to load a pre-trained I3D model and extract the features and save the features as numpy arrays. The charades_dataset_full.py script loads an entire video to extract per-segment features.