Multi-granularity representations for human inter-actions: Pose, motion and intention

Tracking people and their body pose in videos is a central problem in computer vision. Standard tracking representations reason about temporal coherence of detected people and body parts. They have difficulty tracking targets under partial occlusions or rare body poses, where detectors often fail, since the number of training examples is often too small to deal with the exponential variability of such configurations. We propose tracking representations that track and segment people and their body pose in videos by exploiting information at multiple detection and segmentation granularities when available, whole body, parts or point trajectories. Detections and motion estimates provide contradictory information in case of false alarm detections or leaking motion affinities. We consolidate contradictory information via graph steering, an algorithm for simultaneous detection and co-clustering in a two-granularity graph of motion trajectories and detections, that corrects motion leakage between correctly detected objects, while being robust to false alarms or spatially inaccurate detections. We first present a motion segmentation framework that exploits long range motion of point trajectories and large spatial support of image regions. We show resulting video segments adapt to targets under partial occlusions and deformations. Second, we augment motion-based representations with object detection for dealing with motion leakage. We demonstrate how to combine dense optical flow trajectory affinities with repulsions from confident detections to reach a global consensus of detection and tracking in crowded scenes. Third, we study human motion and pose estimation. We segment hard to detect, fast moving body limbs from their surrounding clutter and match them against pose exemplars to detect body pose under fast motion. We employ on-the-fly human body kinematics to improve tracking of body joints under wide deformations. We use motion segmentability of body parts for re-ranking a set of body joint candidate trajectories and jointly infer multi-frame body pose and video segmentation. We show empirically that such multi-granularity tracking representation is worthwhile, obtaining significantly more accurate multi-object tracking and detailed body pose estimation in popular datasets