demianzhang / thumos14-i3d

I3D models trained on Kinetics

Overview

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.

Tensorflow code is from Deepmind's Kinetics-I3D.

Pytorch code is from Kinetics-I3D

Fine-tuning and Feature Extraction

These models were pretrained on imagenet and kinetics (see original repo) for details).

Something to say

You need to down load the checkpoint from the original repo

1. default load the kinetics pre-trained model
2. extract features in thumos14 validation and test dataset the extract way is segment the video at uniform interval.

frames	interval	video fps
<=15000	24	30
<=30000	48	30
>30000	96	30

In order to reduce the redundancy in frames, we choose to subsample the video to 10fps. A clip includes 48 frames, we sample 16 frames and send to the I3D network to extract [1,1024] features

Feature is generated after Mix_5c and avg_pool layer:

input -> output:

rgb: [1, 16, 224, 224, 3] -> [1024,]

flow:[1, 16, 224, 224, 2] -> [1024,]