CMU VASC talk: Detecting Pedestrians by Learning Shapelet Features and Boosted Multiple Deformable Trees for Parsing Human Poses

Greg Mori
Simon Fraser University
**Friday, May 25, 3:30pm**

Detecting Pedestrians by Learning Shapelet Features and Boosted Multiple Deformable Trees for Parsing Human Poses

In this talk we present two pieces of work in the "Looking at People" domain. In the first part, we address the problem of detecting pedestrians in still images. We introduce an algorithm for learning shapelet features, a set of mid-level features. These features are focused on local regions of the image and are built from low-level gradient information that discriminates between pedestrian and non-pedestrian classes. Using AdaBoost, these shapelet features are created as a combination of oriented gradient responses. To train the final classifier, we use AdaBoost for a second time to select a subset of our learned shapelets. By first focusing locally on smaller feature sets, our algorithm attempts to harvest more useful information than by examining all the low-level features together. We present quantitative results demonstrating the effectiveness of our algorithm. In particular, we obtain an error rate 14 percentage points lower (at $10^{-6}$ FPPW) than the previous state of the art detector of Dalal and Triggs on the INRIA dataset.

In the second part, we present a method for estimating human pose in still images. Tree-structured models have been widely used for this problem. While such models allow efficient learning and inference, they fail to capture additional dependencies between body parts, other than kinematic constraints. In this paper, we consider the use of multiple tree models, rather than a single tree model for human pose estimation. Our model can alleviate the limitations of a single tree-structured model by combining information provided across different tree models. The parameters of each individual tree model are trained via standard learning algorithms in a
single tree-structured model. Different tree models are combined in a discriminative fashion by a boosting procedure. We present experimental results showing the improvement of our model over previous approaches on a very challenging dataset.