About me
Hi, I am Madhawa, from the Machine Learning Production Engineering team of Sanctuary AI. Previously, I worked as an SDE II at Microsoft Mixed Reality.
I graduated from Simon Fraser University with an M.Sc. in Computing Science. My thesis project Plan2Scene introduced a novel method to create 3D scenes of interiors using floorplans and photos. This work was supervised by Dr. Manolis Savva, Dr. Angel Chang, and Dr. Yasutaka Furukawa, and published at CVPR 2021. I was attached to the GrUVi Lab of SFU and my studies were supported by a graduate fellowship and a graduate research assistantship.
Previously, I worked as a Software Engineer in the Machine Learning and Data Science team of CodeGen International. I did my bachelor’s in Computer Science and Engineering from the University of Moratuwa, Sri Lanka. In parallel, I performed well in several programming and product innovation competitions such as Google Summer of Code, Microsoft Imagine Cup, and IEEEXtreme.
News
- 2023 March: I joined the Machine Learning Production Engineering team of Sanctuary AI.
- 2021 December: I joined the Microsoft Mixed Reality Team.
- 2021 September: I graduated with an M.Sc. in Computing Science from Simon Fraser University.
- 2021 June: Presented our work Plan2Scene at CVPR 2021.
- 2021 March: My research paper Plan2Scene is accepted to CVPR 2021.
Awards, Grants and Competitions
- SFU Computing Science Graduate Fellowship - Fall 2019
- Best Paper Award Finalist (Top 3): 30th International Conference on Computer Applications in Industry and Engineering (CAINE 2017), California, U.S.A.
- Microsoft Imagine Cup 2017: National Level Finalist (Sri Lanka) For product TravelSphere, a VR based mobile app made for travel enthusiasts.
- Microsoft Imagine Cup 2016 - Word Citizenship Category: National Level Runners Up (Sri Lanka) For developing a VR based solution for Phobia treatments.
- Google Summer of Code 2017 - Certificate of Completion: for project PySiddhi with WSO2.
- IEEEXtreme Programming Competition:
- Year 2017: Global Rank 160 (Top 6.3%) and Country Rank 7 (Top 2.9%)
- Year 2016: Global Rank 197 (Top 9.2%) and Country Rank 8 (Top 2.6%)
- Year 2015: Global Rank 162 (Top 6.5%) and Country Rank 10 (Top 3.8%)
- Year 2014: Global Rank 146 (Top 7.7%) and Country Rank 10 (Top 5.5%)
- University of Moratuwa - IBM Engineering Scholarship for the undergraduate student with highest semester GPA: Semesters 3 and 4.
- University of Moratuwa: Department of CSE award for the undergraduate student with highest semester GPA: Semesters 2 and 3.
- University of Moratuwa: Virtusa award for the student with highest marks for Software Engineering semester project.
Research
Plan2Scene: Converting Floorplans to 3D Scenes
CVPR 2021
Abstract - We address the task of converting a floorplan and a set of associated photos of a residence into a textured 3D mesh model, a task which we call Plan2Scene. Our system 1) lifts a floorplan image to a 3D mesh model; 2) synthesizes surface textures based on the input photos; and 3) infers textures for unobserved surfaces using a graph neural network architecture. To train and evaluate our system we create indoor surface texture datasets, and augment a dataset of floorplans and photos from prior work with rectified surface crops and additional annotations. Our approach handles the challenge of producing tileable textures for dominant surfaces such as floors, walls, and ceilings from a sparse set of unaligned photos that only partially cover the residence. Qualitative and quantitative evaluations show that our system produces realistic 3D interior models, outperforming baseline approaches on a suite of texture quality metrics and as measured by a holistic user study.
Accelerated Human Pose Estimation
February 2019 - September 2019
- Optical-Flow based frame-rate enhancement for arbitrary 2D human pose estimators in a time-delayed real-time fashion. Introduces a pose tracking solution as a by-product.
- Extensive testing on the OpenPose human pose estimator. A reduction in error of 30–35% at practical frame rates. Improvements of up to 50% in resource constrained environments.
On-field testing of the real-time system, processing front office surveillance video of a banking environment.
LIVE-SENSE - Real-time Human Analytics using Multiple Camera Feeds
January 2017 - December 2017 | Undergraduate Research Project - University of Moratuwa
- Generates floor mapped analytics on movements of people in an environment monitored by a surveillance camera network.
- World space analytics via homography. Improved accuracy if multi-view cues are available.
Aggregated and individual level analytics via re-identification and tracking.
Demo Video
Paper: Real-time Pedestrian Mapping from Camera View to World Space
Paper: Open set person re-identification framework on closed set re-id systems
Open-Source Projects
YOLO3-4-Py
April 2018 - Present | Self initiated hobbyist project
A Python wrapper on vanila YOLO 3.0 implementation in Darknet.
- A fast light-weight Python wrapper on YOLO 3.0 Object Detector, which the authors have implemented on Darknet C++ Machine Learning Library
- This project has 500+ stars and 150+ forks on GitHub.
Supports rapid conversion of Numpy Arrays to Darknet compatible formats. (~7% additional overhead on HD resolution).
PySiddhi
May 2017 - August 2017 | Google Summer of Code collaborative project with WSO2
Python wrapper on Siddhi Streaming Analytics Java Library.
- Initiated PySiddhi, a Python wrapper on WSO2 Siddhi Streaming Analytics Java Library.
- Interconnection of two platforms using Java Native Interface and Python Native Interface.
This was a collaborative project with WSO2, under the Google Summer of Code 2017 program.
Publications
Vidanapathirana, M., Wu, Q., Furukawa, Y., Chang, A., and Savva, M. Plan2Scene: Converting Floorplans to 3D scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. [Project Page, Pre-print]
Vidanapathirana, M., Sudasingha, I., Vidanapathirana, J., Kanchana, P. and Perera, I., 2019. Tracking and frame-rate enhancement for real-time 2D human pose estimation. The Visual Computer 36 (2020): 1501 - 1519. [Link, Pre-print]
Vidanapathirana, M., Meegahapola, L.B. and Perera, I.,2017. Cognitive Analysis of 360 degree Surround Photos. in 2017 Future Technologies Conference (FTC). [Link]
Vidanapathirana, M., Sudasingha, I., Kanchana, P., Vidanapathirana, J. and Perera, I., 2017. Open Set Person Re-Identification Framework on Closed Set Re-Id Systems. in 2017 IEEE International Conference on Signal and Image Processing (ICSIP). [Link, Pre-print]
Vidanapathirana, M., Meegahapola, L.B. and Perera, I.,2017. Optimizing Cognitive Analysis Sensitivity of Photospheres using Cube Maps. in 2017 IEEE International Conference on Signal and Image Processing (ICSIP). [Link, Pre-print]
Meegahapola, L.B., Vidanapathirana, M. and Perera, I., 2017. Intelligent Digital Signage Platform for Targeted Advertising. in 30th International Conference on Computer Applications in Industry and Engineering (CAINE 2017). [Pre-print]
Vidanapathirana, M., Sudasingha, I., Kanchana, P., Vidanapathirana, J. and Perera, I., 2017. Real-time Pedestrian Mapping from Camera View to World Space. in 2017 IEEE 12th International Conference on Industrial and Information Systems (ICIIS). [Link, Pre-print]
Hobbyist Projects
Balance IT - Space
2012 - 2013 | 1000+ downloads on Google Play
This started as a pet-project I did to learn Android, OpenGL and Accelerometer sensors. I was able to extend that work to a published game in the Google Play store. Some key learnings from this projects:
- Targetting to a wide device range with different performance and screen specifications. It supported low spec devices such as the Samsung Galaxy Y while providing an improved experience to high-end device users (E.g. Galaxy S3).
- Dealing with lags introduced by intermittent Java GC calls.
- Google Analytics and behavior of PlayStore (It reached 1000+ downloads back in 2014).
Fragment shaders, vertex shaders, and OpenGL ES 2.0.