MobileSAM Annotator

A blazing fast (and simple) annotator tool using SAM (Segment Anything Models) using the awesome MobileSAM.

`mobilesam_dir_single_point.py`

This script take all the images inside a folder and request a single point from the user to segment the desired object (single one per image).

What it looks like?

Requirements:

pip install matplotlib \ 
            git+https://github.com/ChaoningZhang/MobileSAM.git \
            torch \
            opencv-python
            timm

Quickstart:

1. Download the model from MobileSAM

wget https://github.com/ChaoningZhang/MobileSAM/raw/master/weights/mobile_sam.pt -P mobilesam_weights/

2. Run!

Sample exec:

python3 mobilesam_dir_single_point.py --image_dir_path sample_images/
                                        --mobilesam_weights_path mobilesam_weights
                                        --output_folder out_sample_images_segmentation/ 
                                        --resize_width 640

All arguments:

usage: mobilesam_dir_single_point.py [-h] --image_dir_path IMAGE_DIR_PATH
                                     [--mobilesam_weights_path MOBILESAM_WEIGHTS_PATH]
                                     [--exclude_subdirs | --no-exclude_subdirs] [--output_folder OUTPUT_FOLDER]
                                     [--use_image_center_point | --no-use_image_center_point]
                                     [--resize_width RESIZE_WIDTH]

options:
  --image_dir_path IMAGE_DIR_PATH
                Path to the image directory
  --mobilesam_weights_path MOBILESAM_WEIGHTS_PATH
                Path to the MobileSAM weights
  --exclude_subdirs
                Exclude subdirectories in the image directory
  --output_folder OUTPUT_FOLDER
                (Optional) Output folder to save the results (segmentations and masks)
  --use_image_center_point
                Instead of clicking a point, use the center of the image to guide the segmentation
  --resize_width RESIZE_WIDTH
                Resize images, for visualization purposes, to the desired width and original aspect ratio.

If your object is centralized in all images, you don't need to click anything, just use --use_image_center_point

TODO:

Convert masks to bounding boxes annotations, specially in the COCO format;
Decouple interface code;
Make possible to click several times and for new classes;

Acknowledges

MobileSAM made it possible to really use SAM for annotation, since it's incredibly accurate for its speed (it runs at 250ms per image in my laptop!).

gustavofuhr/mobilesam_annotator