Improving Perception in PCL

Stefan Holzer of TU Munich spent last spring working on including new algorithms into the Point Cloud Library (PCL). His main focus was on creating a recognition and machine learning library in PCL. 

With respect to the recognition library, Stefan assisted in implementing LINEMOD [1], a highly efficient template-matching approach for detecting texture-less objects in heavily cluttered scenes. LINEMOD is based on Kinect data. For detecting objects, it uses color gradients computed from image data, as well as surface normals, computed from depth data.

Within the machine learning library, Stefan worked on implementing a flexible and efficient decision tree learning framework. This was applied to obtain keypoint detectors with improved repeatability and efficiency. Additionally, tools were created to evaluate existing detectors and to create the data for learning.

Having a state-of-the-art object detection method in PCL enables efficient detection of texture-less objects with kinect data. Thanks to the new learning framework there is now the ability to create keypoint detectors with improved detection characteristics, which are useful for localization and object detection.

[1] S. Hinterstoisser, S. Holzer, C. Cagniart, S. Ilic, K. Konolige, N. Navab, V. Lepetit Multimodal Templates for Real-Time Detection of Texture-less Objects in Heavily Cluttered Scenes 

IEEE International Conference on Computer