A Simple Detector with Frame Dynamics is a Strong Tracker

[A Simple Detector with Frame Dynamics is a Strong Tracker]
Chenxu Peng, Chenxu Wang, Minrui Zou, Danyang Li, Zhengpeng Yang, Yimian Dai, Ming-Ming Cheng, Xiang Li
In The 4th Anti-UAV Workshop (CVPRW), 2025

News

• 2025/05/07: Our paper was honored with the Best Paper Award at the CVPR Anti-UAV Workshop!
• 2025/04/21: We released our code!
• 2025/03/10: We achieved first place in Track 1 and second place in Track 2 of the 4th Anti-UAV Competition (CVPRW 2025)!

Dataset Preparation

Prepare the raw dataset, or download the manually filtered version with unreliable annotations removed (Password:i285).

/Data/train/
├── 01_1751_0250-1750
│   ├── 000001.jpg
│   │── ...
│   └── IR_label.json
├── 01_2192_0001-1500
│   ├── 000001.jpg
│   │── ...
│   └── IR_label.json
└──...
    ├── 000001.jpg
    │── ...
    └── IR_label.json 

Cross-validation split

Compute similarity between different sequences using image matching algorithms such as SuperPoint and LightGlue.

python ./LightGlue/main.py  # similarity_results.json

Split sequences into training and testing sets based on similarity between sequences.

python ./LightGlue/gen_best_fold.py  # best_folds

Next, perform sampling on the sequences of each fold to obtain the final training and validation sets. The sampling interval can be freely set, such as every 5 or 10 frames.

python ./Data/train_valid_split.py  # best_folds 

Finally, based on the cross-validation splits and the sampling results, write a script to extract the corresponding data and generate the cross-validation sets. Then, convert the data into the MMDetection or YOLO format as needed(Code omitted here).

Methods

1.Frame Dynamics

Determine the target dataset to be used, and convert both its training and validation sets into the required format(Yolo or MMdet format). Select one of the two: generate an optical flow or frame difference dataset for training.

python ./Data/gen_frame_diff.py    # frame difference
# or
python ./Data/gen_optical_flow.py  # optical flow

2.TC-Filtering

TC-Filtering can be applied either during the post-processing stage of object detection inference or at the final model ensembling stage. Here, we only provide the code for model results ensemble. First, the detection results that need to be fused and the data that needs to be inferred should be organized into the given format (./TC_Filtering/results). Then, run the following code according to your own requirements:

python ./TC_Filtering/code_concat.py
python ./TC_Filtering/code_pirorfilter.py
python ./TC_Filtering/code_max_to_1.py
python ./TC_Filtering/code_wbf.py

Installation

conda create -n your_name python==3.10 -y
conda activate your_name
python -m pip install -r requirements.txt

Citation

If you use this in your research, please cite this project.

@article{peng2025a,
	title={A Simple Detector with Frame Dynamics is a Strong Tracker},
	author={Chenxu Peng, Chenxu Wang, Minrui Zou, Danyang Li, Zhengpeng Yang, Yimian Dai, Ming-Ming Cheng, Xiang Li},
	journal={arXiv preprint arXiv:2505.04917},
	year={2025}
}

@article{peng2025a,
  title={A Simple Detector with Frame Dynamics is a Strong Tracker},
  author={Chenxu Peng, Chenxu Wang, Minrui Zou, Danyang Li, Zhengpeng Yang, Yimian Dai, Ming-Ming Cheng, Xiang Li},
  booktitle={CVPRW},
  year={2025}
}