If you failed to install and run this tracker, please email me (dkn2014@mail.dlut.edu.cn)

This is a python-implemented visual object tracking algorithm. Use meta-updater to control the update of ATOM.

• python 3.7
• ubuntu 16.04
• cuda-10.0

1. Clone the GIT repository:

 $ git clone https://github.com/Daikenan/LTMU.git

2. Clone the submodules.
   In the repository directory, run the commands:

   $ git submodule init  
   $ git submodule update

3. Run the install script.

source create -n DiMP_MU python=3.7
source activate DiMP_MU
cd path/to/DiMP_MU
pip install -r requirements.txt

4.Download models Download [metric model] [ATOM model] and put in the following path:

utils/metric_net/metric_model/metric_model.pt
ATOM_MU/pytracking/networks/atom_default.pth

5. Run the demo script to test the tracker:

cd path/to/ATOM_MU
source activate DiMP_MU
python Demo.py

1. Run original tracker and record all results(bbox, response map,...) on LaSOT dateset to get first round training data of meta-updater.You can modify test_tracker.py like this:

import os
from run_tracker import eval_tracking, p_config
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
p = p_config()
p.save_training_data = True
p.tracker = 'ATOM'
p.name = 'atom_test'
eval_tracking('lasot', p=p, mode='all')

2. In meta_updater/tcopt.py, modify tcopts['lstm_train_dir'] and tcopts['train_data_dir'] like this:

tcopts['lstm_train_dir'] = './atom_mu_test_1'
tcopts['train_data_dir'] = '../results/atom_test/lasot/train_data'

tcopts['train_data_dir'] is dir of the training data and tcopts['lstm_train_dir'] is the save path of training models.

3. Run meta_updater/train_meta_updater to train meta-updater.After training, you can evaluate your own meta-updater by modifying test_tracker.py like this:

import os
from run_tracker import eval_tracking, p_config
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
p = p_config()
p.save_training_data = True
p.tracker = 'ATOM_MU'
p.name = 'atom_mu_test_1'
eval_tracking('lasot', p=p, mode='test')
p.save_training_data = False
eval_tracking('tlp', p=p)
eval_tracking('votlt19', p=p)

And then run evaluate.results.py to evaluate results.

For most trackers, one stage of training can achieve a significant performance improvement. If you want to get better performance, you can try multiple stages of training like this:

4. Run tracker with meta-updater and record all results(bbox, response map,...) on LaSOT dateset to get first round training data of meta-updater.You can modify test_tracker.py like this:

import os
from run_tracker import eval_tracking, p_config
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
p = p_config()
p.save_training_data = True
p.tracker = 'ATOM_MU'
p.name = 'atom_mu_test_1'
eval_tracking('lasot', p=p, mode='all')

5. In meta_updater/tcopt.py, modify tcopts['lstm_train_dir'] and tcopts['train_data_dir'] like this:

tcopts['lstm_train_dir'] = './atom_mu_test_2'
tcopts['train_data_dir'] = '../results/'atom_mu_test_1'/lasot/train_data'

tcopts['train_data_dir'] is dir of the training data and tcopts['lstm_train_dir'] is the save path of training models.

6. Run meta_updater/train_meta_updater to train meta-updater.After training, you can evaluate your own meta-updater by modifying test_tracker.py like this:

import os
from run_tracker import eval_tracking, p_config
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
p = p_config()
p.save_training_data = True
p.tracker = 'ATOM_MU'
p.name = 'atom_mu_test_2'
eval_tracking('lasot', p=p, mode='test')
p.save_training_data = False
eval_tracking('tlp', p=p)
eval_tracking('votlt19', p=p)

And then run evaluate.results.py to evaluate results.

Modify the corresponding p.name,tcopts['lstm_train_dir'],tcopts['train_data_dir'] and repeat 4-6;