learning joint top-down and bottom-up strongprocessesstrong for 3d visual.pdfVIP

Learning Joint Top-Down and Bottom-up Processes for 3D Visual Inference Cristian Sminchisescu Atul Kanaujia and Dimitris Metaxas TTI-C Rutgers University crismin@ kanaujia,dnm @ /˜crismin /˜kanaujia,˜dnm Abstract forward, discriminative recognition methods which predict state distributions directly from image features. Despite be- We present an algorithm for jointly learning a consis- ing simple to understand and fast, recognition methods tend tent bidirectional generative-recognition model that com- to assume that the object of interest is segmented and could bines top-down and bottom-up processing for monocular 3d be blind-sighted by the lack of feedback – they cannot self- human motion reconstruction. Learning progresses in alter- asses accuracy. It may be possible to bootstrap them us- native stages of self-training that optimize the probability ing powerful global or part-based 2d human detectors, but of the image evidence: the recognition model is tunned us- one challenge is to make this approach scale to the range of ing samples from the generative model and the generative poses needed for general 3d reconstruction without progres- model is optimized to produce inferences close to the ones sively modeling most of the 3d constraints. This is partly predicted by the current recognition model. At equilibrium, because the space of (even typical) human articulation is the two models are cons