flâneur

Fine-tune Segment-Anything model. Tips & caveats | by Rustem Glue | Jun, 2023 | Medium | Medium

medium.com · 973 words · saved by 1 readers

In this blog post, we will explore the process of fine-tuning SAM (Segment-Anything-Model), an image semantic segmentation model. We will delve into the reasons behind fine-tuning, what are available strategies and I’ll share tips and caveats from my own experiments. SAM is a powerful image semantic segmentation model designed to accurately predict pixel-level masks for a wide range objects within an image. It consists of three parts: There are two main modes that SAM runs on. The first mode is automatic mask generation (AMG) in which the model generates proposals for mask and then tries to segment all of the areas in an image into unlabelled polygons. The other mode is prompt-guided which usually takes bounding boxes or points as prompt inputs along with an image and outputs a polygon for each prompt. Finally, SAM comes in three different flavors: base, large and huge. The huge model is a large 32-block-deep vision transformer with about 636 million parameters. It takes more time to g

Fine-tune Segment-Anything model Rustem Glue 4 min read · Jun 9, 2023 -- 1 Listen Share In this blog post, we will explore the process of fine-tuning SAM (Segment-Anything-Model), an image semantic segmentation model. We will delve into the reasons behind fine-tuning, what are available strategies and I’ll share tips and caveats from my own experiments. SAM is a powerful image semantic segmentation model designed to accurately predict pixel-level masks for a wide range objects within an image. It consists of three parts: Image encoder —a heavy vision transformer backbone that generates image f

saved by

related reading