/CDT-Summer-School-2021

Primary LanguageMATLAB

CDT-Summer-School-2021

Setup

conda create -n detectron2 python=3.8 -y && conda activate detectron2
  • More setup
pip install pyyaml==5.1 && pip install torch==1.8.0+cu101 torchvision==0.9.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html && pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.8/index.html
  • Do this to see if torch installed correctly and cuda is enabled
python - <<EOF
import torch, torchvision
print(torch.__version__, torch.cuda.is_available())
import cv2 as cv
print(cv.__version__)
EOF

expected output:

1.7.1 True
4.5.1
  • Running python webcam.py with dummy image ./input.jpg will give give something like:

just dont forget to hit 0 to close image or you might crash the terminal sesh. 
python detectron2/demo/demo.py --config-file detectron2/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --webcam --confidence-threshold 0.5 --opts MODEL.WEIGHTS detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl

output runs at around 7FPS on RTX 2080.

animated

  • Quick hack to only show Apple class (pretrained system sees apple with ~97% confidence, increase the threshold to this to exclude classes with lower confidence)
python detectron2/demo/demo.py --config-file detectron2/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml --webcam --confidence-threshold 0.97 --opts MODEL.WEIGHTS detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl

animated