MPI-INF Logo
Computer Graphics

Research Interests

  • Computer Vision
  • Machine Learning and Neural Networks
  • Body Pose Estimation

Publications

Single-Shot Multi-Person 3D Body Pose Estimation From Monocular RGB Input
D. Mehta; O. Sotnychenko; F. Mueller; W. Xu; S. Sridhar; G. Pons-Moll; C. Theobalt
arXiv 1712.03453
A single-shot approach to jointly predict 3D body pose of multiple subjects in general scenes without requiring prior bounding-box extraction. Trained on new composited MuCo-3DHP dataset and evaluated on a new recorded multi-person 3D pose benchmark MuPoTS-3DHP.
[paper]


GANerated Hands for Real-Time 3D Hand Tracking from Monocular RGB
F. Mueller; F. Bernard; O. Sotnychenko; D. Mehta; S. Sridhar; D. Casas; C. Theobalt
Computer Vision and Pattern Recognition (CVPR) 2018
A real-time hand tracking approach that only uses monocular RGB input. Trained on synthetic data enhanced using a new geometrically consistent image-to-image translator for unpaired examples.
[paper] [project page]


MonoPerfCap: Human Performance Capture from Monocular Video
W. Xu; A. Chatterjee; M. Zollhoefer; H. Rhodin; D. Mehta; H.P. Seidel; C. Theobalt
ACM Transactions on Graphics (SIGGRAPH) 2018
A marker-less approach for temporally coherent 3D performance capture of humans with general clothing using only RGB video.
[paper] [project page]


Monocular 3D Human Pose Estimation In The Wild Using Improved CNN Supervision
D. Mehta; H. Rhodin; D. Casas; P. Fua; O. Sotnychenko; W. Xu; C. Theobalt
International Conference on 3D Vision (3DV) 2017
In-the-wild 3D body pose estimation from monocular RGB input through a combination of the new MPI-INF-3DHP human pose dataset with an increased scope of augmentation, transfer learning from 2D pose data, as well as CNN regularization and supervision schemes.
[paper] [project page]


Real-time Hand Tracking under Occlusion from an Egocentric RGB-D Sensor
F. Mueller; D. Mehta; O. Sotnychenko S. Sridhar; D. Casas; C. Theobalt
International Conference on Computer Vision (ICCV) 2017
A method for real-time hand tracking under occlusion in cluttered egocentric scenes from a single RGB-D camera. Trained with a new large-scale dataset SynthHands which was captured using a mixed reality approach, and evaluated on a real benchmark dataset EgoDexter which provides annotated fingertip positions.
[paper] [project page]


VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera
D. Mehta; S. Sridhar; O. Sotnychenko; H. Rhodin; M. Shafiei; H.P. Seidel; W. Xu; D. Casas; C. Theobalt
ACM Transactions on Graphics (SIGGRAPH) 2017
Real-time in-the-wild 3D human pose estimation from a single RGB camera.
[paper] [project page]


Deep Shading: Convolutional Neural Networks for Screen-Space Shading
O. Nalbach; E. Arabadzhiyska; D. Mehta; H.P. Seidel; T. Ritschel
Eurographics Symposium on Rendering (EGSR) 2017
Using learned stacked convolutions (a.k.a. CNNs) for screen-space shading effects generated from deferred shading buffers.
[paper] [project page]


Demos, Theses, Short Papers, and Sundry

VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera
D. Mehta; S. Sridhar; O. Sotnychenko; H. Rhodin; F. Mueller; W. Xu; D. Casas; C. Theobalt
Demo at Conference on Computer Vision and Pattern Recognition (CVPR) 2017
Live demo of real-time in-the-wild 3D human pose estimation from a single RGB camera, running on a laptop.
[demo]


Encoding Spatial Context in Local Image Descriptors
D. Mehta
Master's Thesis (Supervisor: Dr. Roland Angst)
Analysis of the implicit relative orientation information captured by Dense SIFT as the key to its efficacy, and proposal of SIFT extensions which use similar relative orientation information without compromising on in-plane rotation invariance.
[thesis] [poster]


Some Projects From a Previous Lifetime
D. Mehta
Small hardware and signal processing projects done in my undergrad days.


Education