Keypoint-Learning

Description

This framework demonstrates the use of a random forest, trained with the method proposed in [1], as a keypoints detector. The framework is composed by three different projects:

• GenerateTrainingSet: implement the training set generation.
• TrainDetector: starting from samples generated with GenerateTrainingSet, train a random forest for keypoints detection (monoscale only).
• TestDetector: demonstrates keypoints extraction (monoscale only).

If you use this code please refer to:

[1] Learning a Descriptor-Specific 3D Keypoint Detector, Samuele Salti, Federico Tombari, Riccardo Spezialetti, Luigi Di Stefano; The IEEE International Conference on Computer Vision (ICCV), 2015, pp. 2318-2326.

[2] Learning to Detect Good 3D Keypoints, Alessio Tonioni, Samuele Salti, Federico Tombari, Riccardo Spezialetti, Luigi Di Stefano; International Journal of Computer Vision (IJCV), 2017.

Usage

GenerateTrainingSet: implementation of training set generation for random forest train describer in[1].

The algorithm requires a set of calibrated 2.5D views of 3d objects (divided into folders: object_name/2.5D_views/), along with two files, groundTruth.txt containing groundtruth matrix (affine transformation from 2.5D views to full 3d model) and overlappingAreas.txt a list of overalapping areas between 2.5D pairs. For details, refer to the examples in: data/example_groundTruth.txt and data/example_overlappingAreas.txt.

For what concern the descriptor to use, the algorithm is customizable by modifying the function computeDescriptorsPerView() in view_manager.hpp.

To increase efficiency, is possible to enable multithreading defining global variables: MULTITHREAD and MULTIVIEW. The required console arguments are the following:

• distance: euclidean distance to accept points with similar descriptor.
• ext: dataset file extension.
• pathDataset: path to folder with 3d object 2.5D views.
• pathTrainingset: path for generated traning set.
• radiusNegative: radius for negative generation.
• radiusNms: radius for non maxima suppression on positive.
• overlap: overlapping threshold between views.

TrainDetector: this sample train and save random forest using features described in [1]. The required console arguments are the following:

• annuli: annuli for features computation.
• bins: bins for features computation.
• pathDataset: path to dataset (same folder used in GenerateTrainingSet)
• pathTrainingData: path for training data-> Positives in: pathTrainingData\Model_Name\positives and Negatives in: pathTrainingData\Model_Name\negatives.
• pathRF: path for Random Forest.
• radiusFeatures: radius for features computation.
• msc: min samples count of Random Forest.
• nameRF: name of YAML file.
• ntrees: number of trees of Random Forest.

TestDetector: example of keypoints detection on point cloud. The required console arguments are the following:

• pathCloud: path to point cloud.
• pathRF: path to trained random forest.
• radiusFeatures: features support.
• radiusNMS: non maxima suppression radius.
• threshold: minimum forest output score to accept a point as keypoint. Value between 0 and 1.

Data

The folder random_forest contains trained random forest for Laser Scanner dataset:

• FPFH-LaserScanner.yaml.gz: random forest trained with FPFH as descriptor.
• SHOT-LaserScanner.yaml.gz: random forest trained with SHOT as descriptor.
• SPINIMAGES-LaserScanner.yaml.gz: random forest trained with SPINIMAGES as descriptor.

The folder point_cloud contains examples of 2.5D views obtained from Laser Scanner dataset.

Dependencies

• Point Cloud Library 1.8.0
• OpenCV 3.2.0
• and all the other libraries necessary the compile the previous ones

The code has been tested on Windows 10 and Microsoft Visual Studio 2015.