This is a downsampled re-implementation of Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation, trained on a single NVIDIA A10G. More details to follow.
To install requirements:
pip install -r requirements.txt
To train the model detailed in the paper, run the following command:
TRAINING=TRUE/FALSE python -m src.model.deeplab
├── LICENSE
├── README.md <- you are here!
├── report <- reproducibility challenge report
├── requirements.txt <- training environment
└── src <- Source code for use in this project.
├── const.py
├── data
│ ├── cityscapes.py <- cityscapes dataloader creation
│ └── common.py <- Dataset-angosting preprocessing routines
├── model
│ ├── aspp.py <- atrous sparse pyramid pooling layer
│ ├── decoder.py <- semantic and instance decoders
│ ├── deeplab.py <- primary trainer
│ ├── encoder.py <- xception-71 backbone
│ ├── heads.py <- semantic, instance center and instance regression heads
│ ├── loss.py <- weighted bootsrapped cross-entropy for semantic head
│ └── metrics.py <- mIOU, AP, PQ
└── coco_tools.py <- pycocotools extensoions