Feature 3d model

[lufixSch]

Base PR for 3D Modeling of the point cloud

• [x] DRAFT

[lole-elol]

Current state tested with more datasets (before ceilPlaneV2 #18 and improved relative Cuboid parameters):

Delivery Area: Good ground plane detection Bad Cuboids

Kicker: Good ground plane Best Cuboids sofar

Pipes: No ground plane Random Cuboids, General bad performance

Terrains: Ground plane found but at wrong Z value (Cuboids are “floating”) Good "Cuboids", because scene does not conatain any real objects (eg.: Furniture)

[lufixSch]

Results for all test datasets with the new relative parameter feature and ceil detection v2. All results were generated with the same parameters.

Delivery Area

Kicker

Pipes

Terrain

All in all improvements for all datasets. More cuboids were detected and fitted closer onto the relevant points.

Some weird boxes at the ceiling of "Terrains". Maybe because of wrong ceiling detection or too many outliers above the ceiling. And ground plane is not at the right z-Position

[lole-elol]

Tests with pointclouds from the 3D reconstruction team show that we first need to rotate the pc so that the presumable floor/ceil is somewhat parallel to the XY-Plane, in order for the rest of the pipeline working correctly.

It seems that not rotating leads to the pipeline breaking completely, since I was only able to get plots for deliveryArea3 The other two failed during ceiling detection.

(points3D_reconstruct_deliveryArea3.txt)

In general, the point clouds look quite good. By rotating deliveryArea1 90° around the X-axis via pctransform we can show that the cuboid and plane detection work with the amount of detail provided. Furthermore, this validates the theory of the 3D reconstruction team that the order in which the images are used to produce the point cloud has a great impact on its quality. I therefore suggest that we should invest more time in finding a way of sorting the images based on similarity, in order to get better point clouds. Since the 3d-model pipeline is at a point where it works "good enough" and we need more optimization before we enter this section of the pipeline.

Note: The point cloud is not parallel enough, eg. some part lies under the floor, but this is just a demo. (points3D_reconstruct_deliveryArea1.txt rotated)

[lufixSch]

Added rotation based on ground plane to the pipeline. I think we could merge this now and add new branches for smaller features