【Team】Adaptive Bayesian Intelligence Team
【Date】2025/July/10(Thursday) 17:00-18:00(JST)
【Speaker】Talk by Dr. Octave Mariotti, University of Edinburgh,Visiting researcher
Title: Unsupervised 3D Vision and Evaluation Biases
Abstract:Typical computer vision systems treat images as flat, 2D data and largely ignore the 3D structure that underlies the world. In this talk, we show how modelling simple geometric constraints into deep learning models can be use to solve otherwise challenging problems without requiring expansive labels. I will focus on two tasks: generic object pose estimation, and semantic correspondence. Along the way, we will also highlight surprising results in commonly used datasets and metrics, exposing strong issues in the evaluations protocols. To alleviate those, we will look at different metric adaptations that more faithfully reflect true 3D-aware performance.
