Hidden Points Removal

Description

This work presents two HPR methods: 1st - from open3d, 2nd - based by mesh generation. There is marked dataset (link) consisting of 3 point clouds, where is every point labeled with a visibility characteristic. Every cloud was processed by HPR methods and the results were compared. In folder points_removal_scripts you can find both of the methods, experiments.py is used to perform these methods and count an accuracy according to marked dataset. Dockerfile is provided to build docker-image as well.

Installation

1. Run this in your terminal:

git clone git@github.com:prime-slam/hidden-points-removal.git

2. Enter the folder:

cd hidden-points-removal

3. Install Docker if you don't have it yet
4. Build docker-image by running the following command:

docker build -t makeitdense .

6. Download dataset
7. Run image by:

docker run --rm -it -v {path_to_visibility_dataset_folder}/:/workspace/dataset makeitdense

8. To test methods on point cloud run:

python3 experiments.py {path_to_point_cloud}

Compare results

Method	accuracy on 1st cloud	accuracy on 2nd cloud	accuracy on 3rd cloud
open3d	0.541	0.54	0.427
mesh-based	0.798	0.763	0.711

Point cloud before being processed

Point cloud after being processed by open3d method

Point cloud after being processed by mesh-based method