• OS: Ubuntu 18.04 with Python Version=3.6

• Device and Packages: using single GTX-1080Ti / RTX-2080Ti with CUDA 10.1 and

torch==1.1.0
opencv-python==4.10.0.*
torchvision==0.2.2*

• Install NVIDIA/apex package for FP16 optimization.

git clone https://github.com/NVIDIA/apex
git checkout 8818ba9
python setup.py install --cuda_ext --cpp_ext

Please make sure the correct installation of pytorch(1.1.0) and apex, otherwise the memory usage may excess 11GB.

• Download fpn50.weight
• The weight file was extracted from the pretrained FPNSSD512 model released from TorchCV: a PyTorch vision library mimics ChainerCV

• Download the VOC2012 Data( VOCtrainval_11-May-2012.tar)

• Download DataMCD-net.zip

Modify the following lines in the configuration file you are going to run, e.g., configs/mcdon.yaml

FPN_pretrained: 'YOUR_PATH_FPN50/fpn50.weight'
texture_base: 'YOUR_PATH_VOC2012/VOCdevkit/VOC2012'
progress_path: 'YOUR_PATH_CHECKPOINT/progress' 
result_path: 'YOUR_PATH_CHECKPOINT/result'
base_dir: 'YOUR_PATH_DATASET/DataMCD-net'

• YOUR_PATH_FPN50: the path for finding the pretrained weight file fpn50.weight

• YOUR_PATH_VOC2012: the path where you extract the downloaded VOCtrainval_11-May-2012.tar

• YOUR_PATH_CHECKPOINT: any desired path for the checkpoints during training

• YOUR_PATH_DATASET: the path where you extract our dataset file

• Train the DON Baseline on 7 classes of objects

  python trainer.py configs/baseline.yaml

• Train our MCDON (MHSCT-12D M=5,N=2) on 7 classes of objects

  python trainer.py configs/mcdon.yaml

python descriptor_evaluation configs/mcdon.yaml

• Pass -tps N to manually specify the restoring N epoch, otherwise we use the tps value in the configuration file
• In the evaluation window,
  • Adjust the proper threshold on the bar
  • Click on either side of the color images to find the matching region in the other side
  • Press s to sample new pairs for comparison, q to exit

• Pretrained weight file of our MCDON using the setting configs/mcdon.yaml

  • Download progress50.tpdata

  • Place the weight file in the path

    YOUR_PATH_CHECKPOINT/progress/mcdon/progress50.tpdata
    

  • Run the visual evaluation

    python descriptor_evaluation.py configs/mcdon.yaml -tps 50