Extremely Dense Point Correspondences using a Learned Feature Descriptor

The video on the left is the video overlay of the SfM results estimated with our proposed dense descriptor. The video on the right is the SfM results using SIFT

This codebase implements the method described in the paper:

Extremely Dense Point Correspondences using a Learned Feature Descriptor

Xingtong Liu, Yiping Zheng, Benjamin Killeen, Masaru Ishii, Gregory D. Hager, Russell H. Taylor, Mathias Unberath

In 2020 Conference on Computer Vision and Pattern Recognition (CVPR)

Please contact Xingtong Liu (xingtongliu@jhu.edu) or Mathias Unberath (unberath@jhu.edu) if you have any questions.

We kindly ask you to cite this paper if the code is used in your own work.

@INPROCEEDINGS{liu2020extremely,
  author={X. {Liu} and Y. {Zheng} and B. {Killeen} and M. {Ishii} and G. D. {Hager} and R. H. {Taylor} and M. {Unberath}},
  booktitle={2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 
  title={Extremely Dense Point Correspondences Using a Learned Feature Descriptor}, 
  year={2020},
  pages={4846-4855},
  doi={10.1109/CVPR42600.2020.00490}}

Instructions

1. Install all necessary python packages: torch, torchvision, opencv-python, numpy, tqdm, pathlib, torchsummary, tensorboardX, albumentations, argparse, pickle, plyfile, pyyaml, datetime, random, shutil, matplotlib, tensorflow.
2. Generate training data from training videos using Structure from Motion (SfM). Please refer to one data example in this storage for data formats. Color images with the format of {:08d}.jpg are extracted from the video sequence where SfM is applied. camer_intrinsics_per_view stores the estimated camera intrinsic matrices for all registered views. In this example, since all images are from the same video sequence, we assume the intrinsic matrices are the same for all images. The first four rows in this file are focal length along x and y direction, and principal point along x and y direction of the camera for the first frame. motion.yaml stores the estimated poses of the camera coordinate system w.r.t. the world coordinate system . selected_indexes stores all frame indexes of the video sequence. structure.ply stores the estimated sparse 3D reconstruction from SfM. undistorted_mask.bmp is a binary mask used to mask out blank regions of the video frames. view_indexes_per_point stores the indexes of the frames that each point in the sparse reconstruction gets triangulated with. The views per point are separated by -1 and the order of the points is the same as that in structure.ply. visible_view_indexes stores the original frame indexes of the registered views where valid camera poses are successfully estimated by SfM. Note that we provide a python script, named colmap_model_converter.py, to convert the COLMAP format to the ones described above. All types of training data described above can be generated from COLMAP running results consisting of cameras.bin, points3D.bin, and images.bin. One example of using colmap_model_converter.py is:

/path/to/python /path/to/colmap_model_converter.py --colmap_exe_path /path/to/COLMAP.bat --sequence_root /path/to/video/sequence

3. Run train.py with proper arguments for dense descriptor learning. Note the images are assumed to have been undistorted. One usage example is:

/path/to/python /path/to/train.py --adjacent_range 1 50 --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --id_range 1 --batch_size 4 --num_workers 4 --num_pre_workers 4 --lr_range 1.0e-4 1.0e-3 --validation_interval 1 --display_interval 20 --rr_weight 1.0 --inlier_percentage 0.99 --training_patient_id 1 --testing_patient_id 1 --validation_patient_id 1 --num_epoch 100 --num_iter 3000 --display_architecture --load_intermediate_data --sampling_size 10 --log_root "/path/to/training/directory" --training_data_root "/path/to/training/data" --feature_length 256 --filter_growth_rate 10 --matching_scale 20.0 --matching_threshold 0.9 --cross_check_distance 5.0 --heatmap_sigma 5.0 --visibility_overlap 20 

4. Add additional arguments --load_trained_model --trained_model_path "/path/to/trained/model" to continue previous training. Run tensorboard to visualize training progress. One example is: tensorboard --logdir="/path/to/training/directory/".
5. Run test.py with proper arguments to evaluate the pair-wise feature matching performance of the learned dense descriptor model. One example is:

/path/to/python /path/to/test.py --adjacent_range 1 50 --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --num_workers 4 --num_pre_workers 4 --inlier_percentage 0.99 --testing_patient_id 1 --load_intermediate_data --visibility_overlap 20
--display_architecture --trained_model_path "/path/to/trained/model" --testing_data_root "/path/to/testing/data" --log_root "/path/to/testing/directory" --feature_length 256 --filter_growth_rate 10 --keypoints_per_iter 3000 --gpu_id 0

6. dense_feature_matching.py can be used to generate pair-wise feature matches for a SfM algorithm to further process on. One usage example is:

/path/to/python /path/to/dense_feature_matching.py --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --batch_size 1 --num_workers 1 --load_intermediate_data --data_root /path/to/video/sfm/data/ --sequence_root /path/to/video/sequence --trained_model_path /path/to/trained/descriptor/model --feature_length 256 --filter_growth_rate 10 --max_feature_detection 3000 --cross_check_distance 3.0 --id_range 1 --gpu_id 0 --temporal_range 30 --test_keypoint_num 200 --residual_threshold 5.0 --octave_layers 8 --contrast_threshold 5e-5 --edge_threshold 100 --sigma 1.1 --skip_interval 5 --min_inlier_ratio 0.2 --hysterisis_factor 0.7

7. Run colmap_database_creation.py to convert the generated feature matches in HDF5 format to SQLite format, named database.db, that is compatible with COLMAP. One example is:

/path/to/python /path/to/colmap_database_creation.py --sequence_root /path/to/video/sequence

8. Run colmap_sparse_reconstruction.py to run mapper in COLMAP for bundle adjustment to generate sparse reconstruction and camera trajectory. One usage example is:

/path/to/python /path/to/colmap_sparse_reconstruction.py --colmap_exe_path /path/to/COLMAP.bat --sequence_root /path/to/video/sequence

9. Run colmap_model_converter.py again as described in step 2 if you want to generate point cloud-video overlays like the GIFs above. Our method can be well integrated into COLMAP pipeline. One example of point cloud overlay with our descriptor and incremental bundle adjustment from COLMAP is as below. The number of points in this reconstruction is more than 210k.

10. Run point_cloud_overlay_generation.py to generate a point cloud-video overlay video. One example is:

/path/to/python /path/to/point_cloud_overlay_generation.py --sequence_root /path/to/video/sequence --display_image --write_video

Pre-trained Models

The pre-trained weights for dense descriptor network are provided here. The zip file also includes monocular depth estimation models that are only relevant to some of the related projects below.

Related Projects

• SAGE: SLAM with Appearance and Geometry Prior for Endoscopy (ICRA 2022)

• Neighborhood Normalization for Robust Geometric Feature Learning (CVPR 2021)

• Reconstructing Sinus Anatomy from Endoscopic Video -- Towards a Radiation-free Approach for Quantitative Longitudinal Assessment (MICCAI 2020)

• Dense Depth Estimation in Monocular Endoscopy with Self-supervised Learning Methods (TMI 2020)