Person Search with Natural Language Description
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
misc person search with natural language description May 10, 2017
README.md
Retrieval.lua
prepro_data.py
train.lua

README.md

Person Search with Natural Language Description

This project aims at searching person using natural language description. Mainly based on our CVPR 2017 paper Person Search with Natural Language Description. The code is modified from the Neuraltalk2 written by Andrej Karpathy.

Installation

This code is written in Lua and requires Torch. See the Torch installation documentation for more details. To run this code, the following packages must be installed:

Data Preparation

  1. Request the dataset from sli [at] ee.cuhk.edu.hk or xiaotong [at] ee.cuhk.edu.hk (academic only).

  2. Save data into a json file that contains a list of image paths, person ID, and raw captions for each image, of the form:

[{"split": "train", "captions": ["A woman is wearing a gray shirt, a pair of brown pants and a pair of shoes.", "She is wearing a dark grey top and light colored pants."], "file_path": "train_query/p11600_s14885.jpg", "id": 11003}]

You can also directly download the json file and save it to "data/reid_raw.json"

  1. Run prepro_data.py as follows:

    python prepro_data.py --input_json data/reid_raw.json --images_root data/imgs --max_length 50 --word_count_threshold 2 --output_json data/reidtalk.json --output_h5 data/reidtalk.h5
    

Testing

  1. Download the pretrained model from snapshot.

  2. Run Retrieval.lua. The result should be around:

    top- 1 = 19.71%
    top- 5 = 43.20%
    top-10 = 55.24%

Training

  1. Download the VGG-16 network pre-trained model from model.

  2. Run train.lua.

Citation

@article{li2017person,
  title={Person search with natural language description},
  author={Li, Shuang and Xiao, Tong and Li, Hongsheng and Zhou, Bolei and Yue, Dayu and Wang, Xiaogang},
  journal={arXiv preprint arXiv:1702.05729},
  year={2017}
}