Action Recognition with CNN

In this project, Convolutional Neural Networks (CNN) are trained to classify images and videos using Pytorch.

Dataset

Used UCF101 data http://crcv.ucf.edu/data/UCF101.php but took only 10 classes of 101 classes. Each clip has 3 frames and each frame is 64 ∗ 64 pixels. The label of clips are in q3_2_data.mat. trLb are labels for training clips and valLb are labels for validation clips.

CNN is first trained to classify actions for each image. Then, CNN is trained using 3D convolution to classify each clip as a video rather than an image

Kaggle Competitions

1. Action recognition by CNN on images - ranked 10 - http://www.kaggle.com/c/cse512springhw3
2. Action recognition by CNN on videos - ranked 32 - http://www.kaggle.com/c/cse512springhw3video