K-Net/contrib/PanopticDeepLab

• Panoptic DeepLab
• • Model Baselines
  • • Cityscapes
  • Environment Installation
  • Dataset Preparation
  • • Cityscapes
  • Training
  • Evaluation
  • Prediction

English | 简体中文

Panoptic DeepLab

The implementation of Panoptic Deeplab algorithm based on PaddlePaddle.

Panoptic DeepLab has comfirmed that the bottem-up algorithm can achieve state-of-the-art effect for the first time。Panoptic DeepLab predicts three ouputs: Semantic Segmentation, Center Prediction and Center Regression。According to the principle of nearest distance, the pixels of instance category are gathered to the center of the instance to obtain the segmentation result. Finally, according to the rules of majority-vote, semantic segmentation results and instance segmentation results are fused to obtain the final panoptic segmentation results.
It achieves segmentation by assigning each pixel to one category or instance.

Model Baselines

Cityscapes

Backbone	Batch Size	Resolution	Training Iters	PQ	SQ	RQ	AP	mIoU	Links
ResNet50_OS32	8	2049x1025	90000	58.35%	80.03%	71.52%	25.80%	79.18%	model | log
ResNet50_OS32	64	1025x513	90000	60.32%	80.56%	73.56%	26.77%	79.67%	model | log

Environment Installation

1. System environment

• PaddlePaddle >= 2.0.0
• Python >= 3.6+
  PaddlePaddle of GPU version is recommended。Please refer to the official website for detailed installation tutorialsPaddlePaddle

2. Download PaddleSeg repository

git clone https://github.com/PaddlePaddle/PaddleSeg

3. Install paddleseg

cd PaddleSeg
pip install -e .

4. Enter PaddleSeg/contrib/PanopticDeepLab directory

cd contrib/PanopticDeepLab

Dataset Preparation

Please put the dataset in data directory under the PaddleSeg/contrib/PanopticDeepLab.

Cityscapes

Go to CityScapes website to download dataset and make is as the following structure:

cityscapes/
|--gtFine/
|  |--train/
|  |  |--aachen/
|  |  |  |--*_color.png, *_instanceIds.png, *_labelIds.png, *_polygons.json,
|  |  |  |--*_labelTrainIds.png
|  |  |  |--...
|  |--val/
|  |--test/
|  |--cityscapes_panoptic_train_trainId.json
|  |--cityscapes_panoptic_train_trainId/
|  |  |-- *_panoptic.png
|  |--cityscapes_panoptic_val_trainId.json
|  |--cityscapes_panoptic_val_trainId/
|  |  |--  *_panoptic.png
|--leftImg8bit/
|  |--train/
|  |--val/
|  |--test/

Install CityscapesScripts

pip install git+https://github.com/mcordts/cityscapesScripts.git

*_panoptic.png Generation command (need to find the createPanopticImgs.py file)：

python /path/to/cityscapesscripts/preparation/createPanopticImgs.py \
        --dataset-folder data/cityscapes/gtFine/ \
        --output-folder data/cityscapes/gtFine/ \
        --use-train-id

Training

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # 根据实际情况进行显卡数量的设置
python -m paddle.distributed.launch train.py \
       --config configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k_lr00005.yml \
       --do_eval \
       --use_vdl \
       --save_interval 5000 \
       --save_dir output

note: Using --do_eval affects training speed and increases GPU memory consumption, turning on and off as you need.

To view more informationg of parameters, run the following command:

python train.py --help

Evaluation

python val.py \
       --config configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k_lr00005.yml \
       --model_path output/iter_90000/model.pdparams

You can download the pretrained model we provide for evaluation directly.

To view more informationg of parameters, run the following command:

python val.py --help

Prediction

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 # Set the number of GPUs as required.
python -m paddle.distributed.launch predict.py \
    --config configs/panoptic_deeplab/panoptic_deeplab_resnet50_os32_cityscapes_1025x513_bs8_90k_lr00005.yml \
    --model_path output/iter_90000/model.pdparams \
    --image_path data/cityscapes/leftImg8bit/val/ \
    --save_dir ./output/result

You can download the pretrained model we provide for prediction directly.

To view more informationg of parameters, run the following command:

python predict.py --help

Panoptic segmentations results:

Semantic segmentations results:

Instance sementations results: