Skip to content

Latest commit

 

History

History
46 lines (35 loc) · 2.55 KB

README.md

File metadata and controls

46 lines (35 loc) · 2.55 KB
Raw

Open Vocabulary Object Detection and Segmentation with Detic and Segment Anything

This repo uses Detic to detect objects based on a text description of each class (i.e. open-vocabulary detection). It then conditions the Segment Anything model with the detected bounding boxes to get segmentation masks.

Setup

First add the following to your bash profile (assuming you have CUDA 11+): export CUDA_PATH=/usr/local/cuda-11.7/

Next, be sure to use python 3.8. If you have a higher version of python, then install 3.8 and use this.

Either run ./setup.sh (make sure the python command uses python 3.8 in this case!) or follow the steps manually.

Usage

Segmenting example images 1.png and 2.png.

source venv/bin/activate
python main.py 1.png -c bottle mug spoon "mug rack" box cpu bowl -d "cuda:0"
python main.py 2.png -c screwdriver "scrubbing brush" -d "cuda:0"

Troubleshooting

  • Segmentation Fault (core dumped) as soon as you run server.py This stems from an issue with the detectron2 installation. Torch and detectron2 are closely linked: you need to make sure you've installed the torch version with the right CUDA extension corresponding to the detectron2 version (as well as your own system setup). Find the correct detectron2 installation command here. Then, find the corresponding torch version compatible withthat and make sure you have that.
  • PIL.Image.LINEAR doesn't exist. If you see something like the below when trying to run server.py: 
      File "/home/nkumar/detic-sam/venv/lib/python3.10/site-packages/detectron2/data/transforms/transform.py", line 46, in ExtentTransform
  def __init__(self, src_rect, output_size, interp=Image.LINEAR, fill=0):
  AttributeError: module 'PIL.Image' has no attribute 'LINEAR'. Did you mean: 'BILINEAR'?
    Then simply edit the offending file to change Image.LINEAR to Image.BILINEAR.

Licenses and Acks

This code is based on prediction_in_wild, which was made by Kishore Pagidi and is licensed under MIT. His repository in turn uses detectron2, Detic and SAM, which are licensed under Apache-2.0.