This is the subject of JoyBuy AI Competition 2017.
The network build with deep residual netowrk from Facebook, and inspired by Deep Networks with Stochastic Depth, adding drop out inside the residual block, to improve the accuracy.
The data set contains 30 videos (1280x720 30fps) of pig, each video indicates a pig, and duration of each average at 90 seconds. The test set contains 3000 jpg images.
##Preprocess The video had been convert into 120x64, and use center crop function to cut off the video into 64X64. the label use for one hot encode. 70% of video frame use for training, and 10% use for validating, and rest of them use for testing.