Investigating the Ambiguity of SAM when Applied to Depth and RGB Images in Bin Picking Applications

Published

September 29, 2025

Author(s)

Marek Franaszek, Pavlo Piliptchak, Prem Rachakonda, Kamel Saidi

Abstract

The Segment Anything Model (SAM) can segment 2D images of novel objects not seen during training, making it particularly useful for robotic bin picking applications that involve diverse manufacturing parts. However, SAM's flexibility comes with inherent ambiguity: when prompted by a single point, it generates three binary masks with varying confidence levels. Our findings indicate that for complex parts that can be visually decomposed into simpler objects, the medium-confidence masks yield more accurate segmentations. In contrast, simpler, non-decomposable parts are more accurately segmented using the highest-confidence masks. This suggests that different mask lists can be sent to the robot's path planning module to accelerate the bin picking process.

Proceedings Title

Proceedings of 11th International Conference on Sensors and Electronic Instrumentation Advances

Conference Dates

September 24-26, 2025

Conference Location

Ponta Delgada, PT

Pub Type

Conferences

Download Paper

Local Download

Keywords

Foundation Models, Segment Anything Model SAM, Robotic Bin Picking

Citation

Franaszek, M. , Piliptchak, P. , Rachakonda, P. and Saidi, K. (2025), Investigating the Ambiguity of SAM when Applied to Depth and RGB Images in Bin Picking Applications, Proceedings of 11th International Conference on Sensors and Electronic Instrumentation Advances , Ponta Delgada, PT, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=960629 (Accessed September 25, 2026)

Additional citation formats