search

facebookresearch / segment-anything

The repository provides code for running inference with the SegmentAnything Model (SAM), links for downloading the trained model checkpoints, and example notebooks that show how to use the model.
Apache License 2.0
46.98k stars 5.56k forks source link

https://segment-anything.com/demo is working fine. But my local scripts/amg.py is not working as well as https://segment-anything.com/demo #725

Open tugsjargal1 opened 6 months ago

tugsjargal1 commented 6 months ago

Dear all,

I am really happy to know this amazing library is existing and you guys created it. Much appreciate. My case is strange. I am going to try to explain my issue on below:

  1. This is my original image that I wanted to segment:

    original-image

  2. My goal is to segment polygon shape where is placed at the center. I use Macbook M1 chip. So, CUDA is not supported in my case.

python scripts/amg.py --device cpu --checkpoint ../checkpoints/sam_vit_h_4b8939.pth --model-type vit_h --output metadata --input original-image.png

Here is my metadata: [ metadata.zip ]

I got below shape at least. Because it is very differenciatable than other shapes. But this one is not what I want to get. 2

This is what I want to get from the image:

original-image-to-achieve
  1. When I upload this original-image.png to https://segment-anything.com/demo# and selected "Everything", seems that "Everything" option is working fine: Screenshot 2024-04-05 at 17 19 30

How I can work segment-anything as https://segment-anything.com/demo# with "Everything" option on my local environment ?

Do you have any suggestions in my case ?

Thanks in advance! Much appreciate any help.

hjoshi95 commented 6 months ago

I feel the model they have on the website is different from the model they have open sourced for the code. I am also currently in the same boat where I am unable to identically reproduce the same results as the demo

krishnaadithya commented 5 months ago

The demo also uses the same vit_h model but the parameters that they use are different and I don't think they have open-sourced it.

I would refer to the following comment , I was able to improve my segment everything results by modifying these two parameters crop_n_layers=1, and crop_n_points_downscale_factor=2,

tugsjargal1 commented 5 months ago

Hi @krishnaadithya, Thanks for your suggestion. I tried by changing parameters crop_n_layers=1, and crop_n_points_downscale_factor=2. But still no luck.

Do you suggest change other parameters ?

Thanks.