Robust Computer Vision
||Robust Computer Vision|
||Simon Kristoffersson Lind <firstname.lastname@example.org>|
||2023-08-14 – 2024-03-01|
Dealing with uncertainty is central when working with robots.
For example, when grasping an object, the object may not be in a fixed location,
so the robot must adapt its trajectory to reach the object.
When performing vision tasks using a camera, however, uncertainty is not accounted for,
and the camera is assumed to produce ideal images.
Naturally, images may be far from ideal in real world scenarios.
Simple lighting changes may be enough to produce sub-optimal images.
The long term goal of this project is to make vision systems more robust,
by making them adaptable.
Adaptability will be accomplished by introducing parameters in the camera pipeline.
To account for uncertainty, confidence will be introduced.
By producing a measure of confidence, the system will be able to identify and adapt to sub-optimal situations.
The current focus in this project is an attempt to model confidence by essentially performing anomaly detection using an Autoencoder (AE) or a Variational Autoencoder (VAE).
AEs, and VAEs have previously been used successfully in the field of anomaly detection,
so the idea is to construct an anomaly detection problem to identify good or bad images in terms of computer vision tasks.