Tracking Persons-of-Interest via Adaptive Discriminative Features (ECCV 2016)

This is the research code for the paper:

Shun Zhang, Yihong Gong, Jia-Bin Huang, Jongwoo Lim, Jinjun Wang, Narendra Ahuja and Ming-Hsuan Yang. "Tracking Persons-of-Interest via Adaptive Discriminative Features", in Proceedings of European Conference on Computer Vision (ECCV), 2016.

We take the T-ara sequence as an example to evaluate our adaptive feature learning approach in this code. Our project website can be found here:

Project page

Citation

If you find the code and pre-trained models useful in your research, please consider citing:

 @inproceedings{Zhang-ECCV-2016,
       author = {Zhang, Shun and Gong, Yihong and Huang, Jia-Bin and Lim, Jongwoo and Wang, Jinjun and Ahuja, Narendra and Yang, Ming-Hsuan},
       title = {Tracking Persons-of-Interest via Adaptive Discriminative Features},
       booktitle = {European Conference on Computer Vision},
       year = {2016},
       pages = {415-433}
   }

System Requirements

• MATLAB (tested with R2014b on 64-bit Linux)
• Caffe

Installation

1. Download and unzip the project code.

2. Install Caffe. Please follow the Caffe installation instructions to install dependencies and then compile Caffe:

   # We call the root directory of the project code `AFL_ROOT`.
   cd $AFL_ROOT/external/caffe-Triplet-New
   make all -j8
   make pycaffe
   make matcaffe
   

3. Download the T-ara images and extract all images into AFL/data/Tara.

4. Download the AlexNet model:

   cd $AFL_ROOT/external/caffe-Triplet-New
   scripts/download_model_binary.py models/bvlc_reference_caffenet
   

5. Download the VGG-Face Model and put it in $AFL_ROOT/external/caffe-Triplet-New/models/VGG. Download the pre-trained face model and put it in $AFL_ROOT/external/caffe-Triplet-New/models/pretrained_web_face.

Usage

Directly run the script run_Tara_example.sh.

Or run the following commands step by step:

1. Mine constraints:

   cd $AFL_ROOT
   # Start MATLAB
   matlab
   >> genTracklet('Tara')
   

2. Learn adaptive discriminative features:

   cd $AFL_ROOT
   sh shell_scripts/Tara/adapt_Triplet.sh
   

3. Extract features:

   sh shell_scripts/Tara/extract_All_Feas.sh
   

4. Perform hierarchical agglomerative clustering algorithm (you can get Fig. 6(a) in our supplementary materials):

   matlab
   >> clustering_tracklets('Tara')
   

5. Perform a simple multi-face tracking:

   matlab
   >> facetracking('Tara')