view article
Figure 3
Quantitative evaluation of SAM architectures and prompting strategies within SMAXI using a slice of CT data of a LEGO figure (De Carlo et al., 2018View full citation) (a-1) Raw image of the LEGO figure (side view) and (a-2, a-3, a-4) visual comparison of a reference ground truth mask against segmentations generated via bounding box and single-point prompts using SAM 1 (ViT-H) for the LEGO figure (front view). (b-1) Computation time required for segmentation across various SAM 1 (ViT-H, ViT-L, ViT-B) and SAM 2.1 (Hiera-L, Hiera-S, Hiera-T) backbones, highlighting the superior inference efficiency of the Hiera architectures. (b-2) Corresponding IoU scores, demonstrating that bounding box prompts consistently yield high fidelity masks (IoU > 0.95) across all architectures. While the box method significantly outperforms single-point prompting in SAM 1 backbones, the advanced Hiera architectures in SAM 2 demonstrate robust convergence, achieving near-perfect IoU scores for both prompting strategies.

Journal logoJOURNAL OF
SYNCHROTRON
RADIATION
ISSN: 1600-5775
Volume 33| Part 5| September 2026|
https://doi.org/10.1107/S1600577526007745