![\"model](\"https://raw.githubusercontent.com/facebookresearch/segment-anything/main/assets/model_diagram.png\")
SAM: segment anything model!
\nThis is a new segmentation model by Meta. It can be used to segment all the objects from a given image.
\n
Here is a result applied to one image from the demo (that you can find here):
\n![\"sam_example\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F172860%2F7db7fa730527714f134373dd1efa2081%2Fsam_vesiuvius.png?generation=1681147493387483&alt=media\")
Looks interesting without even any training. Will try to fine-tune on the training data and share more later. Stay tuned.
", "messages": [ { "id": "2218926", "postDate": "04/12/2023 06:26:39", "content": "SAM: segment anything model!
\nThis is a new segmentation model by Meta. It can be used to segment all the objects from a given image.
\n
Here is a result applied to one image from the demo (that you can find here):
\n
Looks interesting without even any training. Will try to fine-tune on the training data and share more later. Stay tuned.
", "rawMarkdown": "**SAM**: segment anything model!\n\nThis is a new **segmentation** model by Meta. It can be used to segment all the objects from a given image.\n\n\n\n- Main page: https://segment-anything.com/\n- Blog post: https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/\n- Paper: https://arxiv.org/abs/2304.02643\n- Code: https://github.com/facebookresearch/segment-anything\n- Dataset: https://ai.facebook.com/datasets/segment-anything/\n\nHere is a result applied to one image from the demo (that you can find [here](https://segment-anything.com/demo)):\n\n\n\nLooks interesting without even any training. Will try to fine-tune on the training data and share more later. Stay tuned.", "votes": null }, { "id": "2223422", "postDate": "04/16/2023 08:44:59", "content": "![\"sam](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F172860%2F6568959126d1b2a9d1967244150f196a%2Farchitecture_sam.png?generation=1681633900793770&alt=media\")
The SAM paper contains three main contributions:
\nIs your input one of the volume layers? Did you exported as PNG before importing into the demo? I was taking a look at SAM, but thought it was not possible because it was trained on 3 channel RGB images.
", "rawMarkdown": "Is your input one of the volume layers? Did you exported as PNG before importing into the demo? I was taking a look at SAM, but thought it was not possible because it was trained on 3 channel RGB images.", "votes": null }, { "id": "2224789", "postDate": "04/17/2023 16:56:13", "content": "That's a good point to keep in mind. I have done it on the ink image and not a volume. To make it work on a 3D voxels volume, the model has to be adapted. I will share new insights if I can.
", "rawMarkdown": "That's a good point to keep in mind. I have done it on the ink image and not a volume. To make it work on a 3D voxels volume, the model has to be adapted. I will share new insights if I can.", "votes": null }, { "id": "2232535", "postDate": "04/24/2023 12:36:08", "content": "Consider taking a look at this notebook: https://www.kaggle.com/code/skalskip/how-to-use-segment-anything-model-sam
", "rawMarkdown": "Consider taking a look at this notebook: https://www.kaggle.com/code/skalskip/how-to-use-segment-anything-model-sam", "votes": null }, { "id": "2232557", "postDate": "04/24/2023 13:10:37", "content": "Thanks for the link. 👀
", "rawMarkdown": "Thanks for the link. 👀", "votes": null }, { "id": "2272601", "postDate": "05/24/2023 15:48:31", "content": "Did anybody try to fine-tune?
", "rawMarkdown": "Did anybody try to fine-tune?", "votes": null }, { "id": "2272780", "postDate": "05/24/2023 18:27:35", "content": "I am working on something but not yet there.
\nIf you want to have an example notebook, check this one or this one.
Great links, thanks!!
", "rawMarkdown": "Great links, thanks!!", "votes": null } ], "comments": [ { "id": 2223422, "author_name": "yassinealouini", "author_url": "", "post_date": "04/16/2023 08:44:59", "content": "
The SAM paper contains three main contributions:
\nIs your input one of the volume layers? Did you exported as PNG before importing into the demo? I was taking a look at SAM, but thought it was not possible because it was trained on 3 channel RGB images.
", "votes": null, "replies": [ { "id": 2224789, "author_name": "yassinealouini", "author_url": "", "post_date": "04/17/2023 16:56:13", "content": "That's a good point to keep in mind. I have done it on the ink image and not a volume. To make it work on a 3D voxels volume, the model has to be adapted. I will share new insights if I can.
", "votes": null, "replies": [] } ] }, { "id": 2232535, "author_name": "skalskip", "author_url": "", "post_date": "04/24/2023 12:36:08", "content": "Consider taking a look at this notebook: https://www.kaggle.com/code/skalskip/how-to-use-segment-anything-model-sam
", "votes": null, "replies": [ { "id": 2232557, "author_name": "yassinealouini", "author_url": "", "post_date": "04/24/2023 13:10:37", "content": "Thanks for the link. 👀
", "votes": null, "replies": [] } ] }, { "id": 2272601, "author_name": "polishch", "author_url": "", "post_date": "05/24/2023 15:48:31", "content": "Did anybody try to fine-tune?
", "votes": null, "replies": [ { "id": 2272780, "author_name": "yassinealouini", "author_url": "", "post_date": "05/24/2023 18:27:35", "content": "I am working on something but not yet there.
\nIf you want to have an example notebook, check this one or this one.
Great links, thanks!!
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2218926": "**SAM**: segment anything model!\n\nThis is a new **segmentation** model by Meta. It can be used to segment all the objects from a given image.\n\n\n\n- Main page: https://segment-anything.com/\n- Blog post: https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/\n- Paper: https://arxiv.org/abs/2304.02643\n- Code: https://github.com/facebookresearch/segment-anything\n- Dataset: https://ai.facebook.com/datasets/segment-anything/\n\nHere is a result applied to one image from the demo (that you can find [here](https://segment-anything.com/demo)):\n\n\n\nLooks interesting without even any training. Will try to fine-tune on the training data and share more later. Stay tuned.", "2223422": "\n\nThe SAM paper contains three main contributions:\n\n- Promptable segmentation: this tells the model which parts of the image to segment and gives some contextual clues.\n- A general-purpose segmentation model: this is the model part that is made of two encoders, one for the image and one for the prompt and a mask decoder. The image encoder is adapted from the ViT architecture.\n- A data engine: this was used to collect the SA-1B dataset via a feedback loop (model => annotate => data => train => model). The dataset is made of 11M images and more than 1B masks.", "2224752": "Is your input one of the volume layers? Did you exported as PNG before importing into the demo? I was taking a look at SAM, but thought it was not possible because it was trained on 3 channel RGB images.", "2224789": "That's a good point to keep in mind. I have done it on the ink image and not a volume. To make it work on a 3D voxels volume, the model has to be adapted. I will share new insights if I can.", "2232535": "Consider taking a look at this notebook: https://www.kaggle.com/code/skalskip/how-to-use-segment-anything-model-sam", "2232557": "Thanks for the link. 👀", "2272601": "Did anybody try to fine-tune?", "2272780": "I am working on something but not yet there. \nIf you want to have an example notebook, check this [one](https://colab.research.google.com/drive/1F6uRommb3GswcRlPZWpkAQRMVNdVH7Ww?usp=sharing) or this [one](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/SAM/Fine_tune_SAM_(segment_anything)_on_a_custom_dataset.ipynb).", "2273732": "Great links, thanks!!" }, "source": "meta" }