| Summary: | <p>Vision is hard, but the rewards for cracking the problems that stand between us and robust robot vision more than make up for the arduous journey. This work aims to guide the reader through some ideas and concepts that try to make robot vision better under difficult conditions. We tackle the shortcomings of robot vision from three different directions: domain adaptation, image de-noising and finally training data generation. However, the boundaries between these categories are fluid, allowing for complementary operation. </p>
<p>Throughout this work, we successfully demonstrate domain adaptation for localisation across any weather and appearance using generative adversarial networks (GANs), and subsequently extend this to an unsupervised setting that generalises to more than one task. Similarly, we show that convolutional neural networks are a suitable choice for image de-noising - in the context of de-raining and inverse perspective mapping - and present data and ground-truth collection strategies that make this possible. Finally, we tackle the scarcity of training data via data generation. We propose an improvement to cycle-consistency GANs, and a very simple method for collecting images affected by adherent droplets complete with a clear ground-truth. Scene synthesis and randomization techniques are used for improved semantic segmentation models, finishing with an investigation of a collision mitigation system that leverages data created using a simulator.</p>
<p>The techniques described lead to significant gains under difficult conditions such as low illumination, large changes in appearance (domain shift) and noisy data, as well as allowing machine learning models to be trained using data that would normally be very hard, expensive or impossible to obtain.</p>
|
|---|