DCPT

The official implementation for the ICRA2024 paper DCPT: Darkness Clue-Prompted Tracking in Nighttime UAVs.

Models & Raw Results Baidu Driver: 0dof | Training Dataset Labels Baidu Driver: v2nf

Models & Raw Results Google Driver | Training Dataset Labels Google Driver

Framework

News

Jan 29, 2024

DCPT is accepted to ICRA2024 :stuck_out_tongue_closed_eyes: .

Nov 21, 2023

We release codes, models and raw results.

Highlights

:star2: A new unified nighttime tracking framework (no need for enhance then track paradigm).
A gated feature aggregation mechanism is designed for effectively fusing the features between prompters and the foundation model.
DCPT is with high parameter-efficient tuning, containing only 3.03M trainable parameters (~3%).
DCPT achieves SOTA performance on four nighttime benchmarks.

Results

UAVDark135, NAT2021

DarkTrack2021, NAT2021L

Attributed-Based Analysis

Presentation Demo

Usage

Installation

Create and activate a conda environment:

conda create -n DCPT python=3.8
conda activate DCPT

Install the required packages:

bash install.sh

Data Preparation

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

   ${PROJECT_ROOT}
    -- data
        -- bdd100k_night
            |-- images
            |-- annotations
            ...
        -- ExDark
            |-- images
            |-- annotations
        -- shift_night
            |-- 0b3d-e686
            |-- 0b4d-d96f
            ...
        -- trackingnet
            |-- TRAIN_0
            |-- TRAIN_1
            ...
            |-- TRAIN_11
            |-- TEST

Path Setting

Run the following command to set paths:

cd <PATH_of_DCPT>
python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir ./output

You can also modify paths by these two files:

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

Training

Dowmload the pretrained foundation model (OSTrack with Corner Head) and put it under ./pretrained_models/.

python tracking/train.py --script DCPT --config DCPT_Gate --save_dir ./output --mode multiple --nproc_per_node 4 --use_wandb 0

Testing

Download the model weights from Google Drive or BaiduNetDisk

Put the downloaded weights on <PATH_of_DCPT>/output/checkpoints/train/DCPT/DCPT_Gate

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

Some testing examples:

UAVDark135 or other off-line evaluated benchmarks (modify --dataset correspondingly)

python tracking/test.py DCPT DCPT_Gate --dataset uavdark135 --threads 16 --num_gpus 4
python tracking/analysis_results.py # need to modify tracker configs and names

DarkTrack2021

python tracking/test.py DCPT DCPT_Gate --dataset darktrack2021 --threads 16 --num_gpus 4
python tracking/analysis_results.py # need to modify tracker configs and names

NAT2021

python tracking/test.py DCPT DCPT_Gate --dataset nat2021 --threads 16 --num_gpus 4
python tracking/analysis_results.py # need to modify tracker configs and names

NAT2021L

python tracking/test.py DCPT DCPT_Gate --dataset nat2021L --threads 16 --num_gpus 4
python tracking/analysis_results.py # need to modify tracker configs and names

Test FLOPs, and Speed

# Profiling DCPT_Gate
python tracking/profile_model.py --script DCPT --config DCPT_Gate

Acknowledgment

This repo is based on OSTrack and PyTracking library which are excellent works and help us to quickly implement our ideas.

Citation

If our work is useful for your research, please consider citing:

@article{zhu2023dcpt,
  title={DCPT: Darkness Clue-Prompted Tracking in Nighttime UAVs},
  author={Zhu, Jiawen and Tang, Huayi and Cheng, Zhi-Qi and He, Jun-Yan and Luo, Bin and Qiu, Shihao and Li, Shengming and Lu, Huchuan},
  journal={arXiv preprint arXiv:2309.10491},
  year={2023}
}

Contact

If you have any question, feel free to email huayit1@uci.edu.

bearyi26 / DCPT

readme