Gandhimathi (Mathi) Padmanaban
Ph.D., Industrial and Systems Engineering, University of Michigan-Dearborn
Applied Computer Vision and Machine Learning researcher. I build geometry-grounded perception that stays reliable when the camera, platform, or site changes, and that signals when it should not be trusted.
Seeking postdoctoral and research positions, 2026–27. · Download CV
About Me
I am an applied computer vision and machine learning researcher. I completed my Ph.D. in Industrial and Systems Engineering at the University of Michigan-Dearborn, advised by Dr. Fred Feng (dissertation defended June 2026). My work builds perception systems that human-AI systems can rely on, and that keep working when the sensor, platform, or recording site changes. Each feature starts from how a vehicle's projection in the image encodes its geometry, the way its bounding box grows as it approaches and its bearing shifts as it passes, and I derive that quantitatively from the pinhole camera model so the feature carries a physical quantity rather than raw appearance. Geometric-consistency checks reject detections that violate the projection model, and a confidence threshold lets the classification stage abstain rather than commit to a label it cannot support.
I demonstrated this in road-safety measurement, estimating vehicle-bicyclist interactions from a single uncalibrated bicycle camera and showing that the same geometry-grounded formulation transfers to a different platform (the Waymo Open Dataset) when re-fit to a new supervision source. A second line of work models driver behavior from vehicle kinematics, the human side of any system that operates around people. I care about open, reproducible research, and I want to take the same approach, building the real structure of a problem into the model, into new domains where cheap sensors have to make reliable measurements. I am currently seeking postdoctoral and research positions.
Research Focus
I work on perception that has to hold up outside the conditions it was built in. In my dissertation that meant grounding measurements in the geometry of how a camera forms an image, and having the classifier withhold a prediction when its confidence was low. Both were useful, in one domain, and both rested on choices I set by hand: the validation rules, the abstention threshold, and the sensor labels needed to re-fit the model on a new platform.
What interests me now is making those choices less arbitrary. I want to know whether the physics of a scene can drive test-time adaptation to a new platform, in place of the sensor labels my own work relied on, and whether conformal prediction can put a real coverage guarantee behind an abstention threshold I currently set by hand. I am also interested in whether scene geometry can be used to check the spatial claims of large pretrained perception models, which are usually evaluated on clean imagery rather than in the conditions where they get deployed.
News
Education
Research Interests
- Methods: Geometry-informed computer vision • Physics-grounded measurement • Reliability-aware perception (out-of-distribution robustness, confidence-threshold abstention, conformal prediction) • Physics-driven test-time adaptation • Geometric checks on large pretrained perception models • Object detection and tracking • Behavioral modeling from kinematics
- Domains (current and exploring): Infrastructure and platform-agnostic sensing • Trustworthy perception for human-AI systems • Remote sensing and environmental / climate machine learning • Vulnerable road user and transportation safety
Selected Awards
See full CV for the complete list of awards, publications, and service.