LoRAT

The unofficial reproduction for the ECCV 2024 paper Tracking Meets LoRA: Faster Training, Larger Model, Stronger Performance within the pytracking-like framework.

Official Link: https://github.com/LitingLin/LoRAT

[Models] Extraction Code：4EeS

Install the environment

conda create -n lorat python=3.9
conda activate lorat
bash install.sh

Set project paths

Run the following command to set paths for this project

python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir ./output

After running this command, you can also modify paths by editing these two files

lib/train/admin/local.py  # paths about training
lib/test/evaluation/local.py  # paths about testing

Data Preparation

Put the tracking datasets in ./data. It should look like this:

${PROJECT_ROOT}
 -- data
     -- lasot
         |-- airplane
         |-- basketball
         |-- bear
         ...
     -- got10k
         |-- test
         |-- train
         |-- val
     -- coco
         |-- annotations
         |-- images
     -- trackingnet
         |-- TRAIN_0
         |-- TRAIN_1
         ...
         |-- TRAIN_11
         |-- TEST

Training

This repository support the downstream fine-tuning of LoRAT, not the training from the scratch(i.e., we do not implement insert and merge methods of LoRA).

Run:

bash xtrain.sh

Evaluation

Download the model weights from Quark Drive

Put the downloaded weights on $PROJECT_ROOT$/pretrained

Change the corresponding values of lib/test/evaluation/local.py to the actual benchmark saving paths

Run:

bash ytest.sh

Test FLOPs, and Speed

# Profiling base_224
python tracking/profile_model.py --script lorat --config base_224
# Profiling base_378
python tracking/profile_model.py --script lorat --config base_378
# Profiling large_224
python tracking/profile_model.py --script lorat --config large_224
# Profiling large_378
python tracking/profile_model.py --script lorat --config large_378
# Profiling giant_224
python tracking/profile_model.py --script lorat --config giant_224
# Profiling giant_378
python tracking/profile_model.py --script lorat --config giant_378

Performance of Raw Results

Tracker	GOT-10K (AO)	LaSOT (AUC)	TrackingNet (AUC)	LaSOT_Ext(AUC)	TNL2K(AUC)
base_224	72.1	71.7	83.5	50.3	57.3
base_378	73.7	72.9	84.2	53.1	58.4
large_224	75.7	74.2	85.0	52.8	59.5
large_378	77.5	75.1	85.6	56.6	60.7
giant_224	77.7	74.9	85.2	53.3	60.2
giant_378	78.9	76.2	86.0	56.5	61.1