ZhejiangLab Cup Global AI Competition 2018 - Video Recognition and Q&A

Introduction

This is an implement of our model for ZhejiangLab Cup Global AI Competition 2018 - Video Recognition and Q&A.

Contest website: https://tianchi.aliyun.com/competition/introduction.htm?raceId=231676

Final Result

We only ranked 31st in preliminary round and 24th in semifinal round, so we were eliminated (as expected) 😢

Requirement

• One gtx 1080ti is enough to train the model
• CUDA 8.0
• CUDNN6

Installation

pip install -r requirements.txt  # install requirements
imageio_download_bin ffmpeg      # install ffmpeg dependency

Train and Test

./main.py

This will save submit.txt in ../submit folder

You may also need to change data_dir in main.py

Difference performance from different datasets

If only use DatasetA, you will get a result around 0.34.
If you want to achieve the accuracy on ranking board(0.45), you need to merge DatasetA with DatasetB and set the corresponding path in main.py.