Super simple one line of code to use semantic-segment-anything to auto annotate your datasets.

This repo is 100% based on the amazing Semantic-Segment-Anything project but an ultra-simplified version for those who just wanna run one line of code to annotate your datasets and don't give a heck to the underlying processes.

The follows are important

• GPU Memory >= 13 GB (as SAM is GPU memory consuming)
• nvidia docker installed
• nvidia driver installed on host machine

This repo is tested at the following environment

• Ubuntu 22.04
• Single GPU 3090 Ti
• NVIDIA Driver 530

make run

In the input, you will see a demo picture:

whereas the output folder is empty.

Once you are inside the container, run:

make annotate-single-gpu

You will see that the demo image will be annotated and yields results in output folder.

Put all images that you would like to annotate under input, such as:

├── input
│   ├── whatever_name_you_like.jpg
│   ├── whatever_format_you_like.png
│   ├──...

Then:

make annotate-single-gpu

Then, you will get all output in the output folder. That's it.

If you have a really dense objects, the visualization can be hard to tell.

I have prepared one script simple_visualizer.py to parse the json and re-draw the image.

Simply do:

python simple_visualizer.py redraw \
--image_path=input/demo.jpg \
--json_path=output/demo_semantic.json \
--classes_to_keep="[\"person\"]"

And you will see that a new visualization is available: