Designing for Depth Estimation
An ongoing research thread in the Lab for Cognition and Attention in Time and Space (Lab for CATS) is evaluating the behaviour of monocular depth models (depth prediction methods that rely only on a single image). This is a mathematically ill-posed problem, and performing monocular depth estimation relies on a set of heuristics collectively known as monocular depth cues. Like most neural network-based methods, monocular depth models operate largely as black box systems, making it unclear which cues are driving their predictions and how robust they might be to circumstantial changes to cue reliability.
In this project, our lab is attempting to develop new ways of evaluating the behaviour of monocular depth models in order to better understand what depth cues are most important to their predictions and how well these models respond to dynamic changes in cue reliability. In order to explore this, we will be designing and editing scenes in Blender in order to isolate and control specific depth cues, and then analyze the change in model behaviour. I expect that the majority of time this academic year will be devoted to developing and editing Blender scenes, and this will be the primary task of students when initially joining the project.
In your application, please indicate if you are looking to join this semester (Fall 2026) or next semester (Spring 2027).
The Lab for CATS is a collaborative and supportive environment to learn about vision and perception modeling. Students who join the lab get a chance to learn about many of the general aspects of research, while also digging into a specific aspect of vision research.
For this project specifically, students will gain experience with a popular open-source 3D graphics engine (Blender), as well an introduction to a number of conceptual concepts important to spatial vision and depth prediction.