About

The official implementation of the paper

MULTI-TEMPLATE TRACKER DRIVEN BY CACHE MANAGER ALGORITHM, TOWARDS MULTI-DISTRACTOR SCENARIOS

This repository is based on pytracking and Stark. See https://github.com/visionml/pytracking and https://github.com/researchmm/Stark for more help.

Dataset

GOT-10k http://got-10k.aitestunion.com/
LaSOT
TrackingNet
MSCOCO

Environment

python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir .
lib/train/admin/local.py  # paths for training
lib/test/evaluation/local.py  # paths for testing

Reproduce got10k-only version

Train

Download STARK-ST50_got10k_only checkpoint from https://github.com/researchmm/Stark/blob/main/MODEL_ZOO.md

mv %checkpoint% ./checkpoints/train/stark_st2/baseline_got10k_only
# Training step 1
python ./tracking/train.py --script cachet_4_l6 --config baseline_got10k_only --save_dir . --mode multiple --nproc_per_node N --script_prv stark_st2 --config_prv baseline_got10k_only
# Training step 2
python ./tracking/train.py --script cachetcls_4_l6 --config baseline_got10k_only --save_dir . --mode multiple --nproc_per_node N --script_prv cachet_4_l6 --config_prv baseline_got10k_only

Debug

python ./tracking/test.py cache_tracker baseline_got10k_only --dataset DATASET --sequence SEQUENCE --debug 1

Evaluate

python ./tracking/test.py cache_tracker baseline_got10k_only --dataset got10k_test --debug 0 --threads THREADS --num_gpu NUM_GPUS

Pack and submit results

see https://github.com/visionml/pytracking/blob/master/pytracking/util_scripts/pack_got10k_results.py