Celebrity Video Identification Based on Face Features

Introduction

This repository contains codes for 2019 iQIYI Celebrity Video Identification Challenge, which achieved a mAP score of 0.8949 on the test set (Ranked 6th), inspired by Jasonbaby and created by Wenzhe Wang.

Contents

1. Requirements
2. Installation
3. Training
4. Submission
5. Reference

Requirements

1. Python 3.5
2. tensorflow-gpu (I use 1.4.0)
3. Keras (I use 2.0.8)

Installation

1. Clone the iQIYI-VID repository into $VID_ROOT

   git clone https://github.com/zhezheey/iQIYI-VID.git

2. Install python packages you might not have in requirements.txt

   pip install -r requirements.txt

Training

1. Download the IQIYI-VID dataset, then place face_train_v2.pickle and face_val_v2.pickle inside the $VID_ROOT/feat directory, train_gt.txt and val_gt.txt inside the $VID_ROOT/data directory.

2. Train the MLP models (see more details here)

   cd $VID_ROOT/train
   python get_gt.py
   # Change the batch_size in train.py according to your GPU memory.
   sh train.sh

3. By default, trained models are saved under $VID_ROOT/train/model.

Submission

Follow the steps below to build the Docker image of our submission (see more details here).

1. Move the trained models into the $VID_ROOT/docker/resources directory.

2. Build the Docker image

   cd $VID_ROOT/docker
   docker build -t zheey:1.0 -f Dockerfile .

Reference

@article{liu2018iqiyi,
  title={iqiyi-vid: A large dataset for multi-modal person identification},
  author={Liu, Yuanliu and Shi, Peipei and Peng, Bo and Yan, He and Zhou, Yong and Han, Bing and Zheng, Yi and Lin, Chao and Jiang, Jianbin and Fan, Yin and others},
  journal={arXiv preprint arXiv:1811.07548},
  year={2018}
}