Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Static Segmentation by Tracking: A Label-Efficient Approach for Fine-Grained Specimen Image Segmentation

Created by
  • Haebom

Author

Zhenyang Feng, Zihe Wang, Jianyang Gu, Saul Ibaven Bueno, Tomasz Frelek, Advikaa Ramesh, Jingyan Bai, Lemeng Wang, Zanming Huang, Jinsu Yoo, Tai-Yu Pan, Arpita Chowdhury, Michelle Ramirez, Elizabeth G. Campolongo, Matthew J. Thompson, Christopher G. Lawrence, Sydne Record, Neil Rosser, Anuj Karpatne, Daniel Rubenstein, Hilmar Lapp, Charles V. Stewart, Tanya Berger-Wolf, Yu Su, and Wei-Lun Chao.

Outline

This paper studies trait segmentation in the biological domain, especially in specimen images (e.g., butterfly wing stripes, beetle elytra). This detailed task is important for understanding the biology of organisms, but it is labor-intensive as it requires manually annotating segmentation masks for hundreds of images per species. To address this problem, we propose a label-efficient method called static segmentation by tracking (SST), based on the important insight that specimens of the same species exhibit natural variation, but the traits of interest appear consistently. SST concatenates specimen images into “pseudo-videos” and reframes trait segmentation as a tracking problem. SST generates masks for unlabeled images by propagating annotated or predicted masks from “pseudo-before” images. Based on state-of-the-art video segmentation models such as Segment Anything Model 2, SST achieves high-quality trait segmentation with only one labeled image per species, making a breakthrough in specimen image analysis. To further improve the segmentation quality, we introduce a circular consistency loss that again requires only one labeled image. We also demonstrate the broad potential of SST, including one-shot instance segmentation and feature-based image retrieval in natural images.

Takeaways, Limitations

•
Takeaways:
◦
We present a label-efficient method (SST) that achieves high-quality feature segmentation with only one labeled image per species.
◦
Save time and effort over traditional manual annotation tasks.
◦
Effective application of state-of-the-art video segmentation models such as Segment Anything Model 2 to biological image analysis.
◦
It presents potential for a variety of applications, including one-shot instance segmentation and feature-based image retrieval.
◦
Improving segmentation quality through loss of circular consistency.
•
Limitations:
◦
The performance of the proposed method may depend on the quality and consistency of the images used to generate the “pseudo-videos”.
◦
Additional evaluation of generalization performance for different types of biological samples is needed.
◦
Applicability to species showing extreme trait variation needs to be reviewed.
◦
Experimental validation on large datasets may be lacking.
👍
Made with Slashpage