Multi Scale CRF Based RGB-D Image Segmentation Using Inter Frames Potentials Taha Hamedani Robot Perception Lab Ferdowsi University of Mashhad The 2 nd RSI International Conference on Robotics and Mechatronics (ICROM 2014) Oct 2014
Outline Image segmentation Definition Kinect Range Data and its Noises Conditional Random Field (CRF) Region Extraction based on Geometrical Features Energy Function Experimental Result Future Work 2
Image Segmentation Partitioning the image in to the similar and disjoint regions These regions have the similar features such as RGB, orientation, texture 3
Microsoft Kinect Sensor Estimate Depth data by structure light method Project pseudo Random Pattern IR Estimation based on comparison with position of received IR ray 4
2.5 Dimension Data and Noises 5
Conditional Random Field (CRF) According to Hammersley-Clifford Theorem 6
Multi scale CRF Build a Pyramid of variant resolutions of image Compute unary and pairwise potentials for each level of pyramid based on coarsest level of pyramid 7
Energy minimization Solve energy minimization problem as a top down strategy Projection relaxation 8
Inter level Cliques Consider inter level cliques beside singleton and doubleton cliques in each level (Kato) 9
Region Extraction Extract the main region of the scene based on geometrical features RGB edges (canny edge detector) Depth edges (cosine of angle between normals of two neighbor pixels) Sum two edges Morphological opening in order to construct regions of the scene 10
Region Extraction A simple scene with three side wall 11
Region Extraction More complex scene ( NYU V2) 12
Inter Frame Cliques Consider inter frame cliques between current pixel and previous frame labeling 13
Energy Function Define our energy function based on these new cliques and regions 14
Experimental Result Data set New York v2 (NYUV2) Release in 2012 RGB-D images from Kinect 464 images of 26 different indoor scenes Annotated for 1000 classes 15
Experimental Result 16
Experimental Result Hausdorff Distance 17
Future Work Using previous frame data as a more effective manner Information updating after each minimization iteration such as Normal vector correction 18
Question ? Thanks for your attention ? 19