Deep Autoencoder for Combined Human Pose Estimation and Body Model Upscaling and 4D reconstruction
COM2 Level 4
Executive Classroom, COM2-04-02
Abstract:
I will present a method for simultaneously estimating 3D human pose and body shape from a sparse set of wide-baseline camera views. We train a symmetric convolutional autoencoder with a dual loss that enforces learning of a latent representation that encodes skeletal joint positions, and at the same time learns a deep representation of volumetric body shape. We harness the latter to up-scale input volumetric data by a factor of 4x while recovering a 3D estimate of joint positions with equal or greater accuracy than state of the art. Inference runs in real-time (25 fps) and has the potential for passive human behaviour monitoring where there is a requirement for high fidelity estimation of human body shape and pose.
Biodata:
As a Senior lecturer at the University of Surrey, I have a passion for enabling and exploring how to make intelligent machines or computers that can be creative and understand their world. My research ranges from intelligent creative arts such as style aware Inpainting and 4D performance capture. Through to enabling machines to perceive and understand their surroundings with real-time accurate 3D human pose estimation for large-scale markerless motion capture for use both indoors and outdoors, complex, realistic activity recognition and early work on tracking people on vast surveillance networks.
I am interested in employing state of the art technologies using minimal specialist hardware by applying the latest modelling and learning techniques. I see this as a key enabler for future AR and VR experiences, allowing users to create, develop and enjoy immersive technologies through all the devices they own.
I have published over 45 articles in the leading international vision conferences and journals providing state of the art advancements in this field. Most recently I've worked on the InnovateUK projects REFRAME and TotalCapture, leading advances in video-rate actor performance capture for use in film and games production. I'm also a member of the British Machine Vision Association (BMVA) Executive committee and coordinate the national BMVA technical meetings. Bringing together key expert from industry and academia to discuss and identify solutions to current problems in specialist areas of the computer vision and machine learning field.