Master Thesis Presentation: Emil Sandelin
Contact: carl [dot] olsson [at] math [dot] lth [dot] se
Save event to your calendar
Evaluation of Instance Segmentation Methods in the Context of Construction Waste
With the progress made within the field of computer vision the last few years, it is starting to become possible to perform instance segmentation in real time with relatively cheap hardware. This paper's main purpose is to investigate different methods of instance segmentation in the specific case of segmenting construction waste with only a single class. The paper experiments with two models, Mask RCNN and SOLOv2, using ResNet and DenseNet backbones of different depths, and evaluates it according to average precision and prediction time. The author recommends a ResNet50-FPN Mask RCNN model, due to having a good precision-time tradeoff and being easy to implement in the current system.