Recent years have seen plenty of studies that use machine vision as a means of everyday object detection. Transparent objects possess specific properties that make their detection challenging; it follows that fewer machine learning papers address transparent objects. Our work aims to enable the integration of robotics into a real laboratory environment. Therefore we introduce a transparent object detection system. As we base our method on deep learning, the key elements of our solution are data preparation and parameter tuning.
First, the thesis aims to provide an image dataset emphasising transparent objects commonly found within laboratory environments; beakers, Erlenmeyer flasks, microtiter plates, Petri dishes and Florence flasks. We generate multiple datasets versions because we face difficult decisions at the dataset preparation stage, for which we found no preexisting related work. Dataset versions are evaluated in the experimental section of our work.
Following on, we identify key parameters that affect model accuracy. Transparent object detection on synthetic domain datasets proves trivial for our model. Tests on a real domain dataset show the robustness of our model and prove its generalisation abilities. Our model’s performance is at first evaluated with machine vision metrics; primary (m)AP metric, PASCAL VOC (m)AP metric and precise (m)AP metric amount to 89,22 %, 96,48 % and 95,52 %, respectively. We also evaluate model's performance in the robotics context; we estimate that using a gripper, which is open at 115,27 % of inferred object's width, our system would be able to grasp 85,2 % of target transparent objects.
|
