Shreyas Hampali

Email  /  GitHub  /  Google Scholar  /  CV  

Research

I am a Research Scientist at Meta working on hand pose estimation. Prior to that, I completed Ph.D. at TU Graz, Austria , Austria under the supervision of Prof. Vincent Lepetit . I am interested in problems related to hand-object interaction particularly their pose estimation and object reconstruction.

Before starting PhD in 2018, I spent 3 years each at Qualcomm Research, Bangalore and Intel Technologies, Bangalore working on many projects in the area of camera hardware pipeline tuning, image and video processing solutions and their hardware implementations.

• Defended my PhD Thesis on 29/09/2023: thesis
• In-hand 3D object scanning from an RGB Sequence accepted at CVPR'23
• I joined Meta as Research Scientist on Oct'22
• We are organizing HANDS challenge at ECCV'22 workshop.
• Code for 'Keypoint Transformer' is now public.
• 'Keypoint Transformer' accepted as Oral at CVPR 2022. My first ever in-person conference!
• Received sponsorship from Facebook Reality Labs for carrying research on hand-object interaction
• Paper rejected at ICCV!
• Third version of HO3D dataset is out! More accurate annotations, Report, visit website to download
• MCTS for scene understanding accepted at CVPR 2021
• 2 co-authored papers accepted at ECCV 2020
• Honnotate to appear in CVPR 2020

PhD Thesis: 3D Pose and Shape Estimation of Objects and Hands in Challenging Scenarios

We propose solutions for shape and pose estimation of hands and objects in challenging scenarios.

In-Hand 3D Object Scanning from an RGB Sequence.

We propose for the first time a solution for reconsturcting objects in 3D from in-hand manipulation caputered using a monocular RGB camera. We do not assume any constraints on the type of interaction and our method works on diverse set of obejcts.

Keypoint Transformer: Solving Joint Identification in Challenging Hands and Object Interactions for Accurate 3D Pose Estimation.

We propose an efficient network architecture for estimating pose of two hands and object during complex interaction. We also release the challenging H2O-3D dataset, which contains two hands interacting with YCB objects.

Monte Carlo Scene Search for 3D Scene Understanding

We propose a Monte-Carlo Tree Search (MCTS) based analysis-by-synthesis method to recover complete scene (3D layout+objects) from a RGB-D scan of the environment. *Equal contribution

General 3D Room Layout from a Single View by Render-and-Compare.

We propose an analysis-by-synthesis method to estimate a 3D layout of the room - walls, floors, ceilings - from a single perspective view. The method recovers complex non-cubiod layouts by solving a constrained discrete optimization problem.

Measuring Generalisation to Unseen Viewpoints, Articulations, Shapes and Objects for 3D Hand Pose Estimation under Hand-Object Interaction

We measure the generalization capabilities of several approaches during hand pose estimation tasks. Tasks include hand pose estimation from depth maps/RGB images with/without object interaction. Accuracy of the algorithms participating in HANDS 2019 challenge is analysed.

HOnnotate: A method for 3D Annotation of Hand and Object Poses

We propose a method for automatically generation 3D pose annotations for hand and object during interaction. The resulting dataset, HO-3D is publicly available.

Design and source code from Jon Barron's website