This repo was adapted from fb.resnet.torch to the particular problem of this project. The aim is to predict whether a video is running forwards or backwards using a triplet siamese CNN.
The dataset is available here and contains 180 videos of ~200 frames each. It was split into 144 training videos and 36 test videos.
Here is a brief description of the files:
- opts.lua All the options you can set (all have preset values and descriptions)
- checkpoints.lua Handles checkpoints saving and retrieving
- dataloader.lua Multi-threader dataloader
- train.lua Handles training, test, score computation
- main.lua Puts everything together
To start a training, simply run the following command in your favourite shell
th main.lua
To evaluate the latest model you have trained, simply run
th main.lua -testOnly true -resume true
- L. Pickup et al., Seeing the arrow of time, CVPR 2014
- I. Misra, et al., Shuffle and Learn: Unsupervised Learning using Temporal Order Verification, ECCV 2016