yoosan / i3d-tensorflow

Inflated 3D ConvNets for video understanding

I3D-TensorFlow

This repo contains the inflated version of the recently popular ConvNets.

Convert weights from the pretrained model

DeepMind have provided the Inception-v1 inflated 3D model, building upon the sonnet. I slightly modified their code and rewrited the i3d model using the protogenetic tensorflow op. The pretrained weights from kinetics-i3d can be easily migrated to the new model. To botain the weights from kinetics-i3d, execute the following instructions.

$ git clone https://github.com/yoosan/i3d-tensorflow
$ cd i3d-tensorflow
$ git clone https://github.com/deepmind/kinetics-i3d
$ python convert_weights.py

Training the I3D model on UCF101

Now I'm preparing the code for training model on the UCF101 dataset. Using the kinetics pretrained weights, we achive a result of 95.2% top-1 accuracy on split 1 of UCF101 with RGB modality.