Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Antiquities. --- Excavations (Archaeology) --- Excavations (Archaeology). --- Minoans --- Minoans. --- Palace of Knossos (Knossos) --- Palace of Knossos (Knossos). --- Crete (Greece) --- Greece

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This thesis aims to enhance the precision and efficiency of PolyPerception's data analysis in the waste sorting and recycling sector by transitioning from multi-object detection to multi-object segmentation. This shift is critical as it potentially enables more accurate estimations of object weight and conveyor belt density, aligns analytics with industry standards, and improves object similarity predictions through precise embeddings. PolyPerception currently employs highly accurate object detection and semantic classification algorithms based on a predefined taxonomy. This work leverages existing data, bounding boxes, and semantic information to develop detailed instance segmentation and background-foreground segmentation masks. The study evaluates the Segment Anything Model (SAM) against a pre-trained U-Net model, starting with a simpler dataset, containing ground truth masks, and progressing to a more complex and realistic one. Initial findings on the simpler database indicate that while U-Net performs well with specific pre-processing and fine-tuning, SAM's prompt capabilities and exceptional performance, with a mean Intersection over Union (IoU) of 0.918 without fine-tuning and 0.970 after fine-tuning, make it well-suited for the PolyPerception's real data. When applied to the second, realistic database, SAM, prompted with bounding boxes, generated precise instance segmentation masks, even for overlapping objects. Further fine-tuning involved selecting accurate masks from inference results as ground truth annotations, resembling SAM's original training procedure. These accurate instance segmentation masks were also utilised to fine-tune SAM for background-foreground segmentation without prompts. Initially, SAM struggled to locate objects without bounding box prompts. However, fine-tuning significantly improved the model’s performance, resulting in a mean IoU of 0.929 without requiring bounding box prompts. As a result, the model eliminates the need for a separate object detection algorithm for background-foreground segmentation. This work presents a robust approach to instance segmentation with bounding box prompts and background-foreground segmentation without prompts, significantly enhancing PolyPerception's data analysis capabilities.