Presentation on theme: "Leonid Pishchulin Arjun Jain Mykhaylo Andriluka Thorsten Thorm¨ahlen Bernt Schiele Max Planck Institute for Informatics, Saarbr¨ucken, Germany."— Presentation transcript:
Leonid Pishchulin Arjun Jain Mykhaylo Andriluka Thorsten Thorm¨ahlen Bernt Schiele Max Planck Institute for Informatics, Saarbr¨ucken, Germany
Introduction Generation of novel training examples Articulated people detection Articulated pose estimation Articulated pose estimation “in the wild” Conclusion
Recent progress in people detection and articulated pose estimation may be contributed to two key factors. Discriminative learning allows to learn powerful models on a large training corpora robust image features enable to deal with image clutter, occlusions and appearance variation
We use the deformable part model (DPM)  and evaluate its performance on the “Image Parsing” dataset . For training we use training sets from the publicly available datasets: DPM-VOC PASCAL VOC 2009 (VOC)  DPM-IP “Image Parsing”(IP)  DPM-LSP “Leeds Sports Poses” (LSP)dataset  DPM-IP-R and DPM-IP-AR  M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL visual object classes (VOC) challenge.IJCV’10.  P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan.Object detection with discriminatively trained part-based models.PAMI’10.  S. Johnson and M. Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC’10.  D. Ramanan. Learning to parse images of articulated objects. In NIPS’06.
Proposes a new joint model for body pose estimation combining pictorial structures [12,14]model with DPM
We define a new dataset based on the LSP by using the publicly available original non-cropped images. This dataset, in the following denoted as “multi-scale LSP”
Propose a novel method for automatic generation of training examples Evaluate our data generation method for articulated people detection and pose estimation and show that we significantly improve the performance Propose a joint model