Master presentation: Anton Andrée

Handledare: Philipp Hülsdunk, Micropsi Industries
Examinator: Kristian Gustafsson
Opponent: Arbnor Zeqiri och Ellen Sandén

Abstract

In applications of computer vision, a system interprets digital images or video and then typically performs some task such as object detection, classification or segmentation. Today, convolutional neural networks are commonly used for these types of tasks, and performance of state-of-the-art systems have frequently been shown to be on par with that of humans. In certain cases, segmenting out image sections that are irrelevant to the task at hand can be a useful method for preventing data leakage.

In this thesis, we train a convolutional neural network to detect and semantically segment robot grippers in videos showing industrial robots performing various tasks. This is done in two steps, consisting of the base training and the fine-tuning. In the first step, the pre-trained classifier network ResNet50 is further trained on a dataset consisting of images of a wide range of grippers, in order to learn general features. In the subsequent step, the model is fine-tuned on a small number of frames from each video prior to segmentation. The results constitute a proof of concept for the method, showing it to have good potential given more data and computational time.

Presentationen genomförs via Zoom

Zoom-länk: https://gu-se.zoom.us/j/7575269465