This repo is based on Ergys Ristani's DeepCC [code, paper]. This tracker is based on MATLAB.

We added multiple functions for performance and utilities, including our locality-aware setting reported in our CVPR 2019 workshop paper (to be released).

Besides, other dataset support are also added including MOT-16 and AI-City 2019.

For AI-City setup, please download the folder from google drive. Note that the official AI-City 2019 track-1 dataset also has to be downloaded. This folder only act as a incremental package.

The folder we provide contains the re-ID features for demo usage.

Before running, please check that the dataset position in get_opts_aic.m is changed as your setting.

opts.dataset_path    = '~/Data/AIC19';

After that, open up MATLAB at the code root directory, first run get_opts_aic.m to finish the setup. Then, type to run add_gps.m to add gps position to the detections.

To run the demo, please open up MATLAB and run val_aic_ensemble.m. This should give you about 79.7 SCT IDF1 and 78.1 MCT IDF1 on the train set.

For the test set, please run test_aic_ensemble.m. However, the test set result must be uploaded to the AI-City server for online test. To do that, please run prepareMOTChallengeSubmission_aic.m.

If you want to train your own re-ID model, please check our other repo open-reid-tracking.

After training the re-ID model and computing the re-ID features for detection bounding boxes (pre-requisite of tracking), please run the view_appear_score.m file to get your own threshold/norm parameters. NOthe that the experiment directory in view_appear_score.m must be changed accordingly before running.

opts.net.experiment_root = 'experiments/zju_lr001_colorjitter_256_gt_val';

After that, you can replace the old parameters. Remember to change the new feature saving directory in val_aic_ensemble.m or test_aic_ensemble.m, and you should be good to go.