Overview

I have implemented scene action recognitions from UCF101 dataset which can be obtained from here. The dataset being too big. You can download a smaller version of it like UCF50 or UCF11.

Data Preprocessing

I converted videos into frames and took only 32 frames from every video for the training of model. These 16 frames were selected from complete video sequence by skipping frames according to video length.

Model used

convlstm.py

densenet.py

resnet.py

resnext.py

slowfast.py

wide_resnet.py

One of the best model for action recognition Slow Fast Networks for Video Recognition worked best. The implementation of this network in pytorch can be found here.

Training

$ python train.py