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.
AI-City 2019 update
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
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
Train your own re-ID model and run the tracker
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
test_aic_ensemble.m, and you should be good to go.