24th June 2013
CVPR 2013 workshop, Portland Oregon
Computer vision tasks are routinely addressed by building a statistical model that can be learned from annotated data and then be used to perform inference on novel test data. Rich models that express the real world faithfully most often are intractable in that inference and estimation within the model class are hard problems. On the other hand, scalable and efficient discriminative models are based on simplifying assumptions that ignore important physical constraints; as result, for many tasks high performance can only be achieved with extremely large training sets. Therefore, a key problem in computer vision applications is in constructing models expressive enough to solve the task at hand while remaining tractable. Many successful examples where this has been achieved have led to breakthroughs for computer vision, such as for example graphcut-based image segmentation or deformable part models (DPM).
The goal of this workshop is to bring together researchers from the computer vision and machine learning community to discuss all issues related to tractable structured prediction models. In particular,
In all aspects the computer vision community has been at the forefront of developing new ideas, in representation (e.g. perturb-and-MAP, sum-product networks, dense random fields, deep generative models), inference (e.g. dense mean field inference, higher-order factors), and estimation (e.g. direct loss minimization).
|9:05-9:55||invited talk: Danny Tarlow||Designing Loss Functions for Structured Prediction(slides,avi,webm)|
|9:55-10:20||Ben London||Collective Activity Detection using Hinge-loss Markov Random Fields (pdf,slides,avi,webm)|
|10:45-11:35||invited talk: Justin Domke||Reducing CRF training to a series of (possibly nonlinear) logistic regression problems (slides,avi,webm)|
|11:35-12:00||Michael Stark||Modeling Instance Appearance for Recognition - Can We Do Better Than EM? (pdf,slides,avi,webm)|
|14:00-14:50||invited talk: Pedro Felzenszwalb||Contour completion with Fields-of-Patterns (slides,avi,webm)|
|14:50-15:15||Charles Dubout||Accelerated Training of Linear Object Detectors (pdf,slides,avi,webm)|
|15:15-15:55||coffee and tea break|
|15:55-16:20||Xiaoyu Wang||Hierarchical Feature Pooling with Structure Learning: A new method for Pedestrian Detection (slides,avi,webm)|
|16:20-17:10||invited talk: Raquel Urtasun||Efficient learning and inference for holistic scene understanding, (slides,avi,webm)|