MemLoTrack

This repository provides the official implementation of MemLoTrack, a memory-augmented LoRA-based tracker.

• Based on the LoRAT codebase: https://github.com/LitingLin/LoRAT
  We sincerely thank the LoRAT authors for making their excellent codebase publicly available.

• Paper: MemLoTrack (MDPI Sensors): https://www.mdpi.com/1424-8220/25/23/7359

Model weight

 : https://drive.google.com/drive/folders/1cr2yX1UagM48qIuWzvYUSfl2V1CUZuAK?usp=sharing

• antiuav410_retrained_weight.bin : We first pretrained the tracker on standard SOT benchmark datasets (e.g., LaSOT, TrackingNet, COCO 2017), and then fine-tuned it on Anti-UAV410.

• antiuav410_scratch_weight.bin : Model trained from scratch on the Anti-UAV410 dataset

Prerequisites

Environment

Assuming you are working on a python environment with pip installed. We recommend running this in a virtual environment.

system packages (ubuntu)

apt update
apt install -y libturbojpeg

install pytorch

PyTorch version should be >= 2.0.

pip install torch torchvision

extra python packages

pip install -r requirements.txt

Dataset

Download

• Anti-UAV410 (project page): https://github.com/HwangBo94/Anti-UAV410
• Anti-UAV410 (download link): https://drive.google.com/file/d/1zsdazmKS3mHaEZWS2BnqbYHPEcIaH5WR/view

The organized directory should look like as follow :

--AntiUAV410/
	|--test
	|--train
	|--val

Prepare consts.yaml

Copy consts.template.yaml as consts.yaml and fill in the path.

Anti_UAV_410_PATH: '/path/to/antiuav410'

Login to wandb (optional)

Register an account at wandb, then login with the command:

wandb login

Training & Evaluation

Note: Our code performs evaluation automatically when model training is completed.

• Model weight is saved in /path/to/output/run_id/checkpoint/epoch_{last}/model.safetensors.
• Performance metrics can be found on terminal output and wandb dashboard.
• Tracking results are saved in /path/to/output/run_id/eval/epoch_{last}/.

Hyperparameter Configuration

To modify training configurations such as total epochs and other detailed hyperparameters, please edit the following file: config/MemLoTrack/run.yaml

Using run.sh helper script (Linux with NVIDIA GPU only)

The following command executes training and evaluation for the datasets defined in config/_dataset/train.yaml and config/_dataset/test.yaml.

# Train and evaluate MemLoTrack model on all GPUs
./run.sh MemLoTrack dinov2 --output_dir /path/to/output

Train and evaluate on specific GPUs

./run.sh MemLoTrack dinov2 --output_dir /path/to/output --device_ids 0,1,2,3

You can set the default settings, e.g. output_dir, in run.sh.

Specific Dataset Instructions (Anti-UAV 410)

To train or evaluate specifically on the Anti-UAV 410 Dataset, use the following --mixin flags:

• For Train & Evaluation: Add --mixin anti_uav_train --mixin anti_uav_test

• For Evaluation Only: Add --mixin anti_uav_test --mixin evaluation

# Train and evaluate MemLoTrack model on all GPUs
./run.sh MemLoTrack dinov2 --mixin anti_uav_train --mixin anti_uav_test --output_dir /path/to/output

# Only evaluate MemLoTrack model on all GPUs
./run.sh MemLoTrack dinov2 --mixin anti_uav_test --mixin evaluation --output_dir /path/to/output --weight_path /path/to/weight

Note: If you encounter any issues with torch.compile, disable is with --mixin disable_torch_compile.

Note: You can disable wandb logging with --disable_wandb.

Evaluation (Success Rate, Precision, Norm-Precision)

You can run evaluation only with the following command:

# evaluation only, defined in config/_dataset/test.yaml
./run.sh MemLoTrack dinov2 --output_dir /path/to/output --mixin evaluation --weight_path /path/to/weight.bin

# evaluation only, defined in config/MemLoTrack/_mixin/anti_uav_test.yaml
./run.sh MemLoTrack dinov2 --output_dir /path/to/output --mixin evaluation --mixin anti_uav_test --weight_path /path/to/weight.bin

The evaluated datasets are defined in config/MemLoTrack/_mixin/anti_uav_test.yaml.

Results are saved in /path/to/output/run_id/eval/epoch_{last}/, where run_id is the current run ID, and epoch_{last} is the last epoch.

State Accuracy (SA) Evaluation (Anti-UAV410 only)

The bundled evaluation code is based on the official Anti-UAV410 repository.

Dataset evaluation modes

trackit/datasets/SOT/datasets/Anti_UAV_410.py accepts an evaluation_mode seed parameter:

• standard (default): keeps valid target-present frames for the existing AUC, precision, and normalized-precision workflow.
• state_accuracy: keeps every frame in its original order and marks target-absent frames with validity=False. This is required for SA evaluation.

Select the mode in a dataset configuration; no source-code commenting or uncommenting is required:

- name: "Anti_UAV_410"
  type: "SOT"
  splits: "test"
  parameters:
    evaluation_mode: "state_accuracy"

The provided config/MemLoTrack/_mixin/anti_uav_test.yaml already enables this mode. Its SA evaluation source intentionally avoids filters that remove invalid or empty annotations, because those filters would delete target-absent frames and break frame alignment.

Run tracking

./run.sh MemLoTrack dinov2 \
  --output_dir /path/to/output \
  --mixin evaluation \
  --mixin anti_uav_test \
  --weight_path /path/to/weight.bin

The result archive is written below /path/to/output/<run_id>/eval/epoch_<last>/. Extract results.zip; it contains a tracker directory with one <sequence-name>.txt file per Anti-UAV410 sequence. SA mode preserves the official sequence names, so no prefix-removal or file-renaming step is needed.

Calculate SA

From the repository root, run:

python Anti-UAV410/Evaluation_for_SA.py \
  --dataset-path /path/to/Anti-UAV410 \
  --pred-path /path/to/extracted-results/<tracker-directory> \
  --split test \
  --mode 1 \
  --output /path/to/eval_details.txt

--mode 1 means XYWH predictions (the format produced by results.zip); use --mode 2 for XYXY input. The evaluator requires the prediction and ground-truth frame counts to match. In text results, a box with non-positive width or height, including 0 0 0 0, represents a target-absent prediction. For Anti-UAV410 SA-mode exports, annotated target-absent positions are encoded as 0 0 0 0; a prediction/existence length mismatch raises an error instead of being silently truncated or padded.

For the official multi-tracker plots and PDF report, configure dataset_path, result_dir, and tracker paths in Anti-UAV410/Evaluation_for_paper_work.py and Anti-UAV410/utils/trackers.py, then run that script from the Anti-UAV410 directory.

Custom Dataset

This page describes how to create a custom dataset for training and evaluation.

Resumable Checkpointing

Add --mixin resumable to the command line to enable resumable checkpointing. This allows you to resume training from the last saved checkpoint if the training process is interrupted.

./run.sh MemLoTrack dinov2 --output_dir /path/to/output --mixin resumable

Or you can set the default value in run.yaml to:

checkpoint:
  - type: "regular"
    epoch_trigger:
      interval: 10
      last: true
    resumable: true # false --> true
    max_to_keep: 5

Now the training process will save checkpoints every 10 epochs, and the last checkpoint will be saved as recovery.yaml in the checkpoint directory.

Load the last checkpoint by specifying the --resume argument:

./run.sh MemLoTrack dinov2 --output_dir /path/to/output --mixin resumable --resume /path/to/output/run_id/checkpoint/recovery.yaml

Citation

@article{park2025memlotrack, title={MemLoTrack: Enhancing TIR Anti-UAV Tracking with Memory-Integrated Low-Rank Adaptation}, author={Park, Jae Kwan and Han, Ji-Hyeong}, journal={Sensors}, volume={25}, number={23}, pages={7359}, year={2025}, publisher={MDPI} }