The accuracy of point cloud of scanning small object with HoloLens2

Question

I used HoloLens2 to scan a rubic‘s cube from a single perspective. Then I got the colored pointcloud shown in the image below using StreamRecording in Hololens2forCV. Another image shows the RGB Frame. As we can see, there are some noise on the back of the Rubik's cube. These points shouldn't be on the back of the rubik's cube as HoloLens2 can't scan the back side. I want to know why there are so many points on the back of the rubik‘s cube.

enter image description here

enter image description here

enter image description here

Answer 1

We think there are two likely causes that match the specific shape and location of the "noise" pixels:

1. Perspective shifts are caused by merging multiple frames where the camera moves slightly and has different angles on the cubes. Your opened Github issue confirms that multiple depth images were merged to get the point cloud.

2. The depth sensors won't fall off perfectly at the physical edges of the objects. There can be a few pixels of gradual falloff on the boundary that will generate the slope shown in your images.

If you want to eliminate this, you may need to examine the depth images before they're combined into the pixel cloud and do some clean-up pre-processing on the depth images. And also, please keep an eye on that GitHub issue in case having any updates.