Segment Any Object Model (SAOM): Real-to-simulation fine-tuning strategy for multi-class multi-instance segmentation

Authors

Mariia Khan, Edith Cowan UniversityFollow
Yue Qiu
Yuren Cong
Bodo Rosenhahn
Jumana Abu-Khalaf, Edith Cowan UniversityFollow
David Suter, Edith Cowan UniversityFollow

Author Identifier

Mariia Khan: https://orcid.org/0000-0001-6662-4607

Jumana Abu-Khalaf: https://orcid.org/0000-0002-6651-2880

David Suter: https://orcid.org/0000-0001-6306-3023

Document Type

Conference Proceeding

Publication Title

Proceedings - International Conference on Image Processing, ICIP

First Page

582

Last Page

588

Publisher

IEEE

School

School of Science

Comments

Khan, M., Qiu, Y., Cong, Y., Rosenhahn, B., Abu-Khalaf, J., & Suter, D. (2024, October). Segment Any Object Model (SAOM): Real-to-simulation fine-tuning strategy for multi-class multi-instance segmentation. In 2024 IEEE International Conference on Image Processing (ICIP) (pp. 582-588). IEEE. https://doi.org/10.1109/ICIP51287.2024.10647744

Abstract

Multi-class multi-instance segmentation is the task of identifying masks for multiple object classes and multiple instances of the same class within an image. The foundational Segment Anything Model (SAM) is designed for promptable multi-class multi-instance segmentation but tends to output part or sub-part masks in the “everything” mode for various real-world applications. Whole object segmentation masks play a crucial role for indoor scene understanding, especially in robotics applications. We propose a new domain invariant Real-to-Simulation (Real-Sim) fine-tuning strategy for SAM. We use object images and ground truth data collected from Ai2Thor simulator during fine-tuning (real-to-sim). To allow our Segment Any Object Model (SAOM) to work in the “everything” mode, we propose the novel nearest neighbour assignment method, updating point embeddings for each ground-truth mask. SAOM is evaluated on our own dataset collected from Ai2Thor simulator. SAOM significantly improves on SAM, with a 28% increase in mIoU and a 25% increase in mAcc for 54 frequently-seen indoor object classes. Moreover, our Real-to-Simulation fine-tuning strategy demonstrates promising generalization performance in real environments without being trained on the real-world data (sim-to-real). The dataset and the code are available here.

DOI

10.1109/ICIP51287.2024.10647744

Access Rights

subscription content