based on the code from YOLACT: Real-time Instance Segmentation

eval.py lines 411-426, adding the ability to do refinement with GAC
train_GAC_loss.py lines 34-120, copy of train.py with the GAC loss, calaulate the distance maps and changing the data loader
layers/modules/multibox_loss2.py lines 34-120, copy of multibox_loss.py with the added GAC loss
few more minor changes to pass the distance maps to the loss module

• Clone this repository and enter it:

  git clone https://github.com/tomer196/YOLACT-GAC.git
  cd YOLACT-GAC

• Set up the environment using one of the following methods:
  • Set up a Python3 environment (e.g., using virtenv).
  • Install packages:

    # Cython needs to be installed before pycocotools
    pip install Cython==0.29.15
    pip install -r requirments.txt 

• download the COCO dataset and the 2014/2017 annotations. Note that this script will take a while and dump 21gb of files into ./data/coco.

  sh data/scripts/COCO.sh

• To train, grab an imagenet-pretrained model and put it in ./weights.
  • For Resnet50, download resnet50-19c8e357.pth from here.
• Run one of the training commands below.
  • All weights are saved in the ./weights directory by default with the file name <config>_<epoch>_<iter>.pth.

# Trains using the original losses
python train.py --config=yolact_resnet50_config

# Trains using the original losses + GAC loss
python train_GAC_loss.py --config=yolact_resnet50_config

You can download the YOLACT trained models:

To evalute the model, put the corresponding weights file in the ./weights directory and run one of the following commands.

# Quantitatively evaluate the original trained model. 
python eval.py --trained_model=weights/yolact_resnet50_54_800000.pth

# Quantitatively evaluate my trained model with GAC loss. 
python eval.py --trained_model=weights/yolact_resnet50_GAC.pth
# or any train model in weights/

# to evaluate with the GAC refinement add the --GAC flag
python eval.py --trained_model=weights/yolact_resnet50_54_800000.pth --GAC
python eval.py --trained_model=weights/yolact_resnet50_GAC.pth --GAC