Combined Object, Shadow and Illumination Detection in 3D

This is a recently proposed project. In this work we are trying to combine shadow and illumination detection along with object detection in 3D point clouds. Currently we are working on the shadow detection part, and has got some interesting ideas.

Here's some recent updates (as of Dec. 2013). Thanks to PCL for this webgl based point cloud viewer. Press number key "3", "4" or "5" to switch between color schemes.

Here's the original point cloud, taken in our Optics Lab.

Here's the detected plane in polar coordinate. Note that the points lifted from the plate is part of detection results, and as you can see the box on table is not part of the plane.

The following is an initial shadow estimation result. (An estimated light source that produces this shadow is available, but not shown in this point cloud). The results will be updated iteratively to find the best match.

Same could be done for more complicated scenes. This following point cloud is a tape on an optics table. Since the table is shiny and has holes on it, the input data is very noisy. some areas near the tape are entirely missing.

However, we can still get a fairly good initial guess.

Mapping the hit map back to input point cloud, we get the following result.

The second part of initialization is a shadow detection algorithm using color consistency test. Several modifications were made for our framework. Below is (again) the box on table. Shadow points are marked blue.

For current codes, please refer to the project Github:

github.com/shengshuyang/PCLCombinedObjectDetection

Contact Me

Email : shuyangs@usc.edu
Cell : (213)716-0718
Address: EEB416, University of Southern California, CA90089