When: Thursday 27th of March, 1:00pm AEDT
Where: This seminar will be partially presented at the ACFR seminar area, J04 lvl 2 (Rose St Building) and partially online via Zoom. RSVP
Speaker: James Gray
Title: Gradient Consistency: A New Take on Variational Optical Flow and Disparity Estimation
Abstract:
Optical flow and disparity estimation aim to determine a displacement field between images of the same scene. Knowledge-driven approaches to these tasks revolve around brightness constancy assumptions, i.e. objects when displaced do not change in appearance. Variational approaches to optical flow and disparity estimation rely heavily on expressing the brightness constancy assumption as a linearised brightness constancy constraint.
In this seminar, we will identify two key problems with variational approaches to optical flow and disparity estimation. Specifically, the limited range over which the linearisation can be accurately applied and situations where the brightness constancy assumptions are incorrect.
The common approach to addressing the limited valid range of the linearisation is a coarse-to-fine scheme, which relies on a schedule to progressively include image data.
This seminar will discuss a novel Gradient Consistency Model which uses gradient consistency information to assess the validity of the linearisation; this is used to weight the image data so that the data which better adheres to the linearisation is emphasised. Instead of relying on a tuned or learned schedule, the Gradient Consistency Model is self-scheduling, since the weights evolve as the algorithm progresses. We show that the Gradient Consistency Model outperforms standard coarse-to-fine schemes in both rate of convergence and accuracy in the domains of multi-image and multi-scale disparity estimation, optical flow and scene flow.
We also propose the use of the Welsch Loss function to address the situations where the brightness constancy assumptions are not correct. We found that the Welsch Loss function outperforms other robust cost functions, such as the l1 norm in the context of multi-view disparity estimation.
Bio:
James Gray graduated from UNSW with a Bachelor of Electrical Engineering and a Bachelor of Physical Science in 2017. He received the UNSW Faculty of Engineering Dean’s Award in 2014, 2015 and 2016. He then worked as a product development engineer at ResMed until he started a PhD in Computer Vision at UNSW in 2020. After completing his PhD in 2024, he now works at the ARIAM Research Hub as a Postdoctoral Research Associated supervised by Donald Dansereau and collaborates with Trendspek. James’s research interests lie in 3D Reconstruction, multi-view depth estimation, optical flow and scene flow. He has published in highly regarded outlets such as the IEEE International Conference on Image Processing (ICIP) and the IEEE International Workshop on Multimedia Signal Processing (MMSP).