YOLO + SAM tools

Tools for creating polygon annotations to train a YOLO instance segmentation model

The main tool is redoYOLOPolygonsWithSAM.py. It is designed to redo polygon and bbox annotations with Meta's SAM model. For example, maybe you have an instance segmentation model that gives mediocre polygon annotations, or have a model that only gives bboxes, or have a library of mediocre or bbox annotations from Roboflow Universe but what you need are nice polygon annotations... or maybe you just don't feel like creating polygon annotations for everything. This tool will convert any of those sources into polygon annotations using SAM, and then optimize the output to be hand editable.

The tool uses the process:

1. Convert all existing YOLO polygons into bounding boxes (existing bbox annotations will be used as is)
2. Use the bounding boxes with the very mature SAM model to give us new masks
3. Convert those masks BACK AGAIN into YOLO polygon annotations
4. Optimize those polygons so they are not too noisy and can still be hand corrected in roboflow or labelbox etc

INPUT: Yolo bounding box and/or polygon annotations

OUTPUT: All annotations redone using Meta's SAM model

Vertices are cleaned up to allow hand editing in Labelbox/Roboflow/etc after running

To use, install with pip:

pip install ultralytics

pip install visvalingamwyatt //for curve simplifying/optimizing YOLO output

pip install image //post processing needs

pip install transformers

pip install segment-anything

Download:

sam_vit_h_4b8939.pth

You can get it here:
https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth

Point paths:

Go into redoYOLOPolygonsWithSAM.py and set your paths accordingly.

NOTE: Make sure to change the sam_checkpoint path variable as well to point to where you downloaded sam_vit_h_4b8939.pth