Recovering human body configurations using pairwise constraints between parts

摘要: The goal of this work is to recover human body configurations from static images. Without assuming a priori knowledge scale, pose or appearance, problem extremely challenging and demands the use all possible sources information. We develop framework which can incorporate arbitrary pairwise constraints between parts, such as scale compatibility, relative position, symmetry clothing smooth contour connections parts. detect candidate parts bottom-up using parallelism, various configuration assemble them together into configurations. To find most probable configuration, we solve an integer quadratic programming with standard technique linear approximations. Approximate IQP allows us much more information than traditional dynamic remains computationally efficient. 15 hand-labeled images are used train low-level part detector learn constraints. show test results on variety