NCTR: Neighborhood Consensus Transformer for feature matching

Repository of NCTR

Due to the space limitation of the paper, we post the appendix and some additional experiment results here.

Requirements

• Python 3 >= 3.5
• PyTorch (>=1.2.0)
• PyTorch Geometric (>=1.5.0)
• OpenCV >= 3.4
• Matplotlib >= 3.1
• NumPy >= 1.18

The requirements are listed in the requirement.txt file. To install all dependencies, run:

pip install -r requirement.txt

Data preparation

R1M dataset

1. Download R1M dataset from official link.

2. To split train/val/test dataset, run

   python train_val_test_r1m.py --input_dir xxx -- output_dir xxx
   

   This produces a validation set and a test set of 500 images, and the rest of the images are the training set.

3. Then generate random homography for validation and test sets for evaluation.

   python get_perspective.py --image_dir xxx
   

Training

We put the settings in the configs/config.yaml file, which user can edit or use unchanged. To launch train , run:

python train_NCTR.py --config_path='configs/config.yaml'

The logging results will be visible inside a log folder + experiment name, specified in config.yaml.

Qualitative Examples

Homography Estimation

Outdoor Pose Estimation