FastPose - Realtime Pose Estimation and Tracking

FastPose

Challenges

• Occlusion, intense light adn rare poses.
• Unconstrained videos such as camera motion, blur, view variants.

Previous method address single prelocated person (eplout pictorial structures model, following deep channels)

Most used Method

Two-states method are most used in this problems. Firstly detect, crop person from the image, then perform the single person pose estimation in the croped patches.

Issues

• Two-stages methods are scale invariant. The second stage only focuses on the keypoint detection task on a fixed scale.
• These methods cannot perform in real-time because of complex procedures (detection, cropping and scaling images, pose estimation).
• ...

FastPose

Aims

• Peform Pose estimation.
• Tracking towards realtime speed.

Stack

• Multi-Task Network (MTN): jointly optimizes Human detection, pose estimation and person Re-ID simutaneously.
• Scale-Normalized paradigm is proposed to alleviate the scale variation problem for the multi-task network.
• Occlusion-aware Re-ID strategy is designed for articulateed multi-person pose tracking in video.

Main Contribution

• With Re-ID feature, design and end-to-end MTN (Multi-task Network)
• Propose new paradigm (Scale-normalized image and feature pyramid) to alleviating scale variation problem (bottleneck of unified top-down method in Pose Estimation).
• Base on image-pyramidm ignore estremely small or large objects to mkake sizes of objects uniformly distributed in the exact range.
• Combine feature pyramid networks (FPN) with scale distribution -> help the network to avoid multi-scale testing.
• Pose information (output from above MTN) utilized to inder occlusion state and achieve the occlusion-aware Re-ID -> Dramatically reduct the ID switches during tracking.