Framework for plane detection algorithms approbation
This framework has pipeline for plane segmentation in point clouds and some visualization tools
What is already done:
- Pipeline for mapping rgb images planes annotations to point cloud and their visualization
- Plane outliers detection and removing using Open3D RANSAC implementation
- Basic plane detector algorithm based on Open3D RANSAC
- Metrics for plane detection: IoU, Dice, classic ones
This framework can map annotated rgb images and their depth component to the point cloud.
Run it with python main.py [dataset_path] --frame_number=[frame_number] --loader=[loader_name] [--annotations_path=[annotations_path] [--disable_annotation_filter_outliers]] [--algo=[plane_detection_algo_name] [--metric=[metric_name_1] --metric=[metric_name_2] ...]], where dataset_path ---
path to the dataset folder, frame_number --- number of the frame in dataset (enumeration started from 0),
loader_name --- name of the dataset loader (tum and icl_tum are available), annotations_path --- path to annotations.xml file,
plane_detection_algo_name --- algorithm to use for detection of planes and metric_name_X --- name of the metric to benchmark the chosen algorithm.
python main.py C:\dataset --frame_num=0 --loader=icl_tum --annotations_path=C:\annotations.xml
| Original depth image | Annotated RGB image |
|---|---|
 |
 |
Framework build point cloud based on depth image and map annotations to it using different colors.
| Result of mapping |
|---|
 |
As you can see, some objects aren't well annotated: lamp and plant on the left side, for example. To fix such mistakes in annotation outlier detector can be used. This framework has one based on RANSAC algorithm.
| Result of mapping with outliers extraction |
|---|
 |
As you can see now, plant and lamp are marked with black --- as not annotated objects, but not as a wall behind them!