Benchmarking the Robustness of UAV Tracking Against Common Corruptions

Welcome

This repository presents a benchmark study focused on the robustness of UAV tracking methods against various common corruptions. Paper link

How to use this repository

Prepare clean dataset

1. Download Datasets:

   • Obtain the DTB70 and UAV123_10fps datasets, which serve as clean foundation for applying corruption.

2. Dataset Setup:

   • Place the downloaded datasets, it's recommended to organize in a structured manner, such as:

     /data/DTB70
     /data/UAV123_10fps
     

Requirements

1. This project is tested with python 3.6+
2. Install the required Python packages:

   pip install -r requirements.txt
   

   If requirements.txt is not provided, manually install the libraries using:

   pip install numpy torch matplotlib PIL imageio opencv-python imgaug scipy
   

3. For ImageNet-C corruptions, follow the additional setup instructions provided in ImageNet-C.

Generating Depth Maps

1. Run Depth Map Generator:
   • Execute the depth map generation script provided in the runs/ directory. Adjust the script's parameters if necessary to point to your dataset locations.
   • Command for generate depth map for DTB70:

     python main.py --phase 'depth_estimation' --seq_dir '/yourpath/DTB70'

   • Command for generate depth map for UAV123_10fps:

     python main.py --phase 'depth_estimation' --seq_dir '/yourpath/UAV123_10fps'

Download libaray Corruptions from ImageNet-C

• Command for install libaray from ImageNet-C:

git clone https://github.com/hendrycks/robustness.git
cd robustness/ImageNet-C/imagenet_c/imagenet_c
cd ..
pip install -e .

Applying Corruptions

1. Select Corruption Types:

   • This benchmark includes 18 corruption types, each with three severity levels(1 3 5). 'runs/generator' by default generate 18 types of corruption together, you could modify corruption types(line 109) if necessary.

2. Generate Corruptions:

   • Example command for DTB70:

     python main.py --seq_dir /data/DTB70 --output_dir /data/UAV-C

   • Example command for UAV123_10fps:

     python main.py --seq_dir /data/UAV123_10fps --output_dir /data/UAV-C

   • Example output dataset structure of UAV-C

     UAV-C/
        ├── Animal1/
        │   ├── contrast/
        │   │   ├── 1/
        │   │   ├── 3/
        │   │   └── 5/
     

Quantitative Results

We evaluated 12 trackers on the UAV-C benchmark. The table below summarizes the performance of each tracker across all types of corruptions. The last row is measured with the metric ( mS_{cor} ), averaged over all corruption types.