Franaszek, M.
, Piliptchak, P.
and Saidi, K.
(2026),
Dependence of Segment Anything Model (SAM) Performance on Part Geometry in Bin Picking Applications, XII Int. Conference on Sensors and Electronic Instrumentation Advances SEIA'26, Granada, ES, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=962316 (Accessed September 25, 2026)
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
Secure .gov websites use HTTPS
A lock (
) or https:// means you’ve safely connected to the .gov website. Share sensitive information only on official, secure websites.
Dependence of Segment Anything Model (SAM) Performance on Part Geometry in Bin Picking Applications
Published
Author(s)
, ,
Abstract
The Segment Anything Model (SAM) can segment novel, previously unseen objects from 2D images. This capability is especially important in automated bin picking applications, where low-volume and high-variability scenarios are common. When prompted by a point, the SAM outputs three masks with varying confidence scores. Prior research indicates that SAM performance depends on the type of parts in a bin: for parts that are visually decomposable into subcomponents, correctly segmented masks are more often associated with the medium confidence score. Conversely, for parts that are not readily visually decomposable into subcomponents, correct masks are typically associated with the highest confidence score. We investigated the transition from visually decomposable to non-decomposable part categories in simulation by using computer aided design (CAD) models with gradually altered geometric features. We found that this transition occurs within a relatively narrow range of feature modifications. This demonstrates the strong sensitivity of the SAM to the specific types of parts within a bin. Consequently, prior knowledge of a part's complexity can be leveraged to better tune the processing of the output of the SAM.Proceedings Title
XII Int. Conference on Sensors and Electronic Instrumentation Advances SEIA'26
Conference Dates
September 30-October 2, 2026
Conference Location
Granada, ES
Pub Type
Conferences
Download Paper
Keywords
Foundation Models, Segment Anything Model, SAM, Robotic Bin Picking
Citation
Issues
If you have any questions about this publication or are having problems accessing it, please contact [email protected].
Created September 24, 2026