Skip to content

Commit

Permalink
Update README.md
Browse files Browse the repository at this point in the history
  • Loading branch information
carlosuc3m authored Dec 9, 2024
1 parent 03611ab commit f89d976
Showing 1 changed file with 5 additions and 0 deletions.
5 changes: 5 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -18,3 +18,8 @@ The inly thing you need to incorporate SAMJ to your software is to add the depen
</dependency>
```

# Available SAM variants
The original SAM is a heavy and computationally expensive model. Its original version requires a GPU as it might take up to several minutes in standard CPUs and might completely freeze lower end hardware. In order to make SAMJ supported by any hardware, we decided to rely in variants of SAM that are light enough to be runnable on any kind of CPU but still keep most of the performance. These variants are:
* [**SAM-2**](https://github.com/facebookresearch/sam2) (tiny, small and large): newest and improved version of SAM released by Meta. Not only this version reduces a lot the computation cost, it also improves the performance. To do so, this model uses a new variant of Vision Transofrmer (ViT), the [Hiera ViT](https://github.com/facebookresearch/hiera).
* [**EfficientSAM**](https://github.com/yformer/EfficientSAM): small Vision Transformer trained by distilling the original SAM Huge transformer using Masked Autoencoders.
* [**EfficientViTSAM**](https://github.com/mit-han-lab/efficientvit): SAM like model that uses a special lightweigth image encoder, EfficientViT.

0 comments on commit f89d976

Please sign in to comment.