This project involves the identification of different actions from video clips where the action may or may not be performed throughout the entire duration of the video. This is done using two CNN models which are 3D-CNN and LSTM models.
Classes: Data is classified into 5 classes {Diving, Jumping, Basketball, Tennis, Walking}.
Train Data Contains 474 videos. Divided into: Class Count Diving 113 Jumping 100 Basketball 89 Tennis 105 Walking 67
Different Shapes of Videos: Shape Count (320.0, 240.0, 239.0) 44 (320.0, 240.0, 201.0) 32 (320.0, 240.0, 151.0) 15 (320.0, 240.0, 238.0) 12 (320.0, 240.0, 105.0) 9 .. .. (320.0, 240.0, 163.0) 1 (320.0, 240.0, 310.0) 1 (320.0, 240.0, 88.0) 1 (320.0, 240.0, 401.0) 1 (320.0, 240.0, 71.0) 1
Test Data Contains 126 videos. Different Shapes of Videos: Shape Count (320.0, 240.0, 239.0) 11 (320.0, 240.0, 179.0) 4 (320.0, 240.0, 101.0) 4 (320.0, 240.0, 119.0) 4 (320.0, 240.0, 115.0) 3 .. (320.0, 214.0, 177.0) 1 (320.0, 240.0, 176.0) 1 (320.0, 240.0, 201.0) 1 (320.0, 240.0, 130.0) 1 (320.0, 240.0, 87.0) 1
A Single Stream Network - One Network for Spatial information. Transfer Learning Model trained on Sports Data. Input Shape: (16, 112, 112, 3)
A Single Stream Network - One Network for Spatial information. Transfer Learning Model trained on ImageNet data. Input Shape: (16, 112, 112, 3)