My talk will include two parts. The first one is on bin-picking robot system based on stereo vision. I will introduce two running systems I developed so far in my company. One is for picking metallic parts. By modeling the part as the two level rectangle object, we use the highlight on the surface as the salient features to reliably and efficiently determine the location and orientation of each object. The other is on a general-purpose picking system. we use a primitive-based 3-view stereo vision to reconstruct 3-D information on the surface of objects. Based on the object model pervious input, the object can be identified. At the same time, the reliability of each result is evaluated. The final picking object in a bin is determined according to the accuracy of result, the surrounding complex etc.
The second part is on spatio-temporal segmentation from a long image sequence of human being. Based on an unified MDL approach, we conduct spatial segmentation of human body from two consecutive images, tracking each part, temporal segmentation and refining spatial segmentation in a long image sequence.
I will also show some videos on these works.