Octave Mariotti, Robert Mitchell, Brian Seipp
Thu 03 Jun 2021, 13:00 - 14:00
Online (Blackboard collaborate)

If you have a question about this talk, please contact: Jodie Cameron (jcamero9)

Octave Mariotti

Title: Unsupervised viewpoint estimation with conditional generation

Abstract: Understanding the 3D world without supervision is currently a major challenge in computer vision as the annotations required to supervise deep networks for tasks in this domain are expensive to obtain on a large scale. In this talk, we address the problem of unsupervised viewpoint estimation formulated as a self-supervised learning task, where image reconstruction from raw images provides the supervision needed to predict camera viewpoint. After explaining the method, we will present some results as well as discuss the limitations and future developments.

 

Robert Mitchell

Title: Decoding the Ultimate Compass: Modelling Cue Integration in Insect Orientation

Abstract: Animal brains are constantly bombarded with noisy information and they need some way of merging this information to generate a stable perceptual model of the world.  Humans are thought to integrate different cues in a Bayesian optimal fashion, but it is difficult to verify such models against any concrete neural pathways. Recent work has exposed cue integration capabilities in insect behaviour, specifically within the insect neural compass. We will discuss a biologically plausible neural circuit for multimodal compass cue integration, a vector-sum heuristic which can be tested in a simple simulation scenario, and how this heuristic performs against traditional models of cue integration in the context of current behavioural data in dung beetles.

 

 

Brian Seipp

Title: Optimized Low Dimensional Feature Exploration through Tactile Sensing

Abstract: During this talk I will explore our solution to learning maximal information about a one-dimensional feature, the boundary between two materials, using only a tactile sensor.  I will discuss both the phantom environment created for simulations, our method to introduce a prior to inform candidate sampling locations, and our ongoing work to use reinforcement learning to develop an automated sampling strategy.