Detect by Track: Making Detector-Free Matcher Trackable

Official PyTorch implementation of Detect by Track: Making Detector-Free Matcher Trackable.

Installation

conda env create -f environment.yaml
conda activate det
pip install -r requirements.txt
pip install mamba-ssm==2.0.3

Checkpoint

Place the DeT checkpoint at weights/jamma.ckpt. The checkpoint format follows JamMa.

Demo

python demo/demo_det.py

The default demo uses the first three Piazza San Marco images in assets/phototourism_sample_images and writes results to demo/output_det.

Evaluation

python -m eval.eval_imc --ckpt_path weights/jamma.ckpt
python -m eval.eval_megadepth --scene_name 0015 --ckpt_path weights/jamma.ckpt
python -m eval.eval_megadepth --scene_name 0022 --ckpt_path weights/jamma.ckpt

The evaluation scripts run NN-JamMa and DeT-JamMa by default. Add a new method by registering one pair-matching function in eval/eval_utils.py.

Acknowledgements

This codebase is based on JamMa. Parts of the upstream code are derived from LoFTR and XoFTR.

Citation

@inproceedings{det2026,
  author    = {Yusuke Sekikawa and Hideki Shirai and Ruka Eto and Yuzhe Hao and Kengo Mitsui and Nakamasa Inoue},
  title     = {Detect by Track: Making Detector-Free Matcher Trackable},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}