/action_recognition

Solving UCF-101 with fastai2

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

Action Recognition

Implementation of various architectures to solve the UCF 101 actions dataset

It is based on the implementations found on Action Recognition.

I try to keep with updated architectures that come out. Right now transformers are all we need... Follow @lucidrains to get the next attention based model ASAP.

Install

First install fastai:

$ pip install fastcore fastai

Results

Results are computed on a random splut 80%/20%. Using fastai2 built-in fit_one_cycle training.

  • train baseline: Implements a Basic Resnet 34 encoder coupled with a simple attention layer over the frames. (91% accuracy)
  • train convlstm: resnet34 encoder + LSTM layer over image features. (84.8% accuracy)
  • train_transformer: Added the new TimeSformer and STAM from @lucidrains implementations.

This package also provides function to download nad process the video dataset into multiple frames.