Big part of this code is used from dbolya/yolact Colab run method is used from yolact-with-google-colab

• YOLACT is a state of the art, real-time, single shot object segmentation algorithm detailed in these papers: YOLACT: Real-time Instance Segmentation YOLACT++: Better Real-time Instance Segmentation

Go to Runtime > Change Runtime Type
Choose GPU (TPU won't work)

• Install required packets

  #Cython needs to be installed before pycocotools
  !pip install cython
  !pip install opencv-python pillow pycocotools matplotlib

• Downgrade torch to accommodate DCNv2

  !pip install torchvision==0.5.0
  !pip install torch==1.4.0

• Clone YOLACT from github

  %cd /content
  # Clone the repo
  !git clone https://github.com/harsul/SiamYolact.git

• DCNv2

  # Change to the right directory
  %cd /content/yolact/external/DCNv2
  # Build DCNv2
  !python setup.py build develop

• Pretrained Weights

  %cd /content
  # Clone the repo
  !git clone https://github.com/chentinghao/download_google_drive.git
  # Create a new directory for the pre-trained weights
  !mkdir -p /content/yolact/weights
  # Download the file
  !python ./download_google_drive/download_gdrive.py 1ZPu1YR2UzGHQD0o1rEqy-j5bmEm3lbyP ./yolact/weights/yolact_plus_resnet50_54_800000.pth

• Make Folders for input and output videos and upload mot chellenge video into input_videos folder

  !mkdir /content/input_videos
  !mkdir /content/output_videos

• Change input and output paths

  %cd /content
  input_path="/content/input_videos/mot17-11.webm"
  output_path = "/content/output_videos/output.mp4"
  !python ./yolact/eval.py --trained_model=./yolact/weights/yolact_plus_resnet50_54_800000.pth --score_threshold=0.15 --top_k=15 -- video_multiframe=4 --video={input_path}:{output_path}