English | 简体中文

News

• [2023-04-11] PaddleSeg v2.8 is released! Check more details in Release Notes.
• Release Segment Anything Model based on PaddlePaddle. Demos are provided to demonstrate the function of automatic full-image segmentation and specified object segmentation with prompt input.
• Release PP-MobileSeg, a lightweight semantic segmentation model for mobile devices. Comparing PP-MobileSeg with other models on the ADE20K dataset, the segmentation accuracy is improved by 1.5%, the inference speed is accelerated by 42.3%, and the number of parameters is decreased by 34.9%.
• Release QualityInspector v0.5, a full-process solution for industrial quality inspection. It provides a unified and configurable pipeline for single-task and multi-task models, integrates detection and segmentation model libraries, and supports three unsupervised quality inspection methods.
• Release PanopticSeg v0.5, a universal panoptic segmentation solution. It provides the full-process capabilities of panoptic segmentation, integrates two models, and has flexible secondary development capabilities.
• [2022-11-30] PaddleSeg v2.7 released a real-time human matting model PP-MattingV2, a 3D medical image segmentation solution MedicalSegV2, and a real-time semantic segmentation model RTFormer.
• [2022-07-20] PaddleSeg v2.6 released a real-time human segmentation SOTA solution PP-HumanSegV2, a stable-version semi-automatic segmentation annotation tool EISeg v1.0, a pseudo label pre-training method PSSL, and the source code of PP-MattingV1.
• [2022-04-20] PaddleSeg v2.5 released a real-time semantic segmentation model PP-LiteSeg, a trimap-free image matting model PP-MattingV1, and an easy-to-use solution for 3D medical image segmentation MedicalSegV1.
• [2022-01-20] We release PaddleSeg v2.4 with EISeg v0.4, and PP-HumanSegV1 including an open-sourced dataset PP-HumanSeg14K.

Introduction

PaddleSeg is an end-to-end high-efficent development toolkit for image segmentation based on PaddlePaddle, which helps both developers and researchers in the whole process of designing segmentation models, training models, optimizing performance and inference speed, and deploying models. A lot of well-trained models and various real-world applications in both industry and academia help users conveniently build hands-on experiences in image segmentation.

Features

• High-Performance Model: Following the state of the art segmentation methods and using high-performance backbone networks, we provide 45+ models and 150+ high-quality pre-training models, which are better than other open-source implementations.

• High Efficiency: PaddleSeg provides multi-process asynchronous I/O, multi-card parallel training, evaluation, and other acceleration strategies, combined with the memory optimization function of the PaddlePaddle, which can greatly reduce the training overhead of the segmentation model, all these allowing developers to train image segmentation models more efficiently and at a lower cost.

• Modular Design: We build PaddleSeg with the modular design philosophy. Therefore, based on actual application scenarios, developers can assemble diversified training configurations with data augmentation strategies, segmentation models, backbone networks, loss functions, and other different components to meet different performance and accuracy requirements.

• Complete Flow: PaddleSeg supports image labeling, model designing, model training, model compression, and model deployment. With the help of PaddleSeg, developers can easily finish all tasks in the entire workflow.

Community

• If you have any questions, suggestions or feature requests, please do not hesitate to create an issue in GitHub Issues.
• Please scan the following QR code to join PaddleSeg WeChat group to communicate with us:

Overview

Industrial Segmentation Models

High Accuracy Semantic Segmentation Models

These models have good performance and costly inference time, so they are designed for GPU and Jetson devices.

Model	Backbone	Cityscapes mIoU(%)	V100 TRT Inference Speed(FPS)	Config File
FCN	HRNet_W18	78.97	24.43	yml
FCN	HRNet_W48	80.70	10.16	yml
DeepLabV3	ResNet50_OS8	79.90	4.56	yml
DeepLabV3	ResNet101_OS8	80.85	3.2	yml
DeepLabV3	ResNet50_OS8	80.36	6.58	yml
DeepLabV3	ResNet101_OS8	81.10	3.94	yml
OCRNet	HRNet_w18	80.67	13.26	yml
OCRNet	HRNet_w48	82.15	6.17	yml
CCNet	ResNet101_OS8	80.95	3.24	yml

Note that:

• We test the inference speed on Nvidia GPU V100. We use PaddleInference Python API with TensorRT enabled. The data type is FP32, and the shape of input tensor is 1x3x1024x2048.

Lightweight Semantic Segmentation Models

The segmentation accuracy and inference speed of these models are medium. They can be deployed on GPU, X86 CPU and ARM CPU.

Model	Backbone	Cityscapes mIoU(%)	V100 TRT Inference Speed(FPS)	Snapdragon 855 Inference Speed(FPS)	Config File
PP-LiteSeg	STDC1	77.04	69.82	17.22	yml
PP-LiteSeg	STDC2	79.04	54.53	11.75	yml
BiSeNetV1	-	75.19	14.67	1.53	yml
BiSeNetV2	-	73.19	61.83	13.67	yml
STDCSeg	STDC1	74.74	62.24	14.51	yml
STDCSeg	STDC2	77.60	51.15	10.95	yml
DDRNet_23	-	79.85	42.64	7.68	yml
HarDNet	-	79.03	30.3	5.44	yml
SFNet	ResNet18_OS8	78.72	10.72	-	yml

Note that:

• We test the inference speed on Nvidia GPU V100. We use PaddleInference Python API with TensorRT enabled. The data type is FP32, and the shape of input tensor is 1x3x1024x2048.
• We test the inference speed on Snapdragon 855. We use PaddleLite CPP API with 1 thread, and the shape of input tensor is 1x3x256x256.

Super Lightweight Semantic Segmentation Models

These super lightweight semantic segmentation models are designed for X86 CPU and ARM CPU.

Model	Backbone	ADE20K mIoU(%)	Snapdragon 855 Inference latency(ms)	params(M)	Links
TopFormer-Base	TopTransformer-Base	38.28	480.6	5.13	config
PP-MobileSeg-Base	StrideFormer-Base	41.57	265.5	5.62	config
TopFormer-Tiny	TopTransformer-Tiny	32.46	490.3	1.41	config
PP-MobileSeg-Tiny	StrideFormer-Tiny	36.39	215.3	1.61	config

Note that:

• We test the inference speed on Snapdragon 855. We use PaddleLite CPP API with 1 thread, and the shape of input tensor is 1x3x512x512. We test the latency with the final argmax operator on.

Model	Backbone	Cityscapes mIoU(%)	V100 TRT Inference Speed(FPS)	Snapdragon 855 Inference Speed(FPS)	Config File
MobileSeg	MobileNetV2	73.94	67.57	27.01	yml
MobileSeg	MobileNetV3	73.47	67.39	32.90	yml
MobileSeg	Lite_HRNet_18	70.75	10.5	13.05	yml
MobileSeg	ShuffleNetV2_x1_0	69.46	37.09	39.61	yml
MobileSeg	GhostNet_x1_0	71.88	35.58	38.74	yml

Note that:

• We test the inference speed on Nvidia GPU V100. We use PaddleInference Python API with TensorRT enabled. The data type is FP32, and the shape of input tensor is 1x3x1024x2048.
• We test the inference speed on Snapdragon 855. We use PaddleLite CPP API with 1 thread, and the shape of input tensor is 1x3x256x256.

Tutorials

Introductory Tutorials

• Installation
• Quick Start
• A 20 minutes Blitz to Learn PaddleSeg
• Model Zoo

Basic Tutorials

• Data Preparation

  • Prepare Public Dataset
  • Prepare Customized Dataset
  • Label Data with EISeg

• Config Preparation

• Model Training

• Model Evaluation

• Model Prediction

• Model Export

  • Export Inference Model
  • Export ONNX Model

• Model Deployment

  • FastDeploy
  • Paddle Inference (Python)
  • Paddle Inference (C++)
  • Paddle Lite
  • Paddle Serving
  • Paddle JS
  • Benchmark

Advanced Tutorials

• Training Tricks

• Model Compression

  • Quantization
  • Distillation
  • Pruning

• FAQ

Welcome to Contribute

• API Documention

• Advanced Development

  • Detailed Configuration File
  • Create Your Own Model

• Pull Request

  • PR Tutorial
  • PR Style

Special Features

• Interactive Segmentation
• Image Matting
• PP-HumanSeg
• 3D Medical Segmentation
• Cityscapes SOTA
• Panoptic Segmentation
• CVPR Champion Solution
• Domain Adaptation

Industrial Tutorial Examples

• Using PP-HumanSegV2 for Human Segmentation
• Using PP-HumanSegV1 for Human Segmentation
• Using PP-LiteSeg for Road Segmentation
• Using PaddleSeg for Face Parsing and Makeup
• Using PaddleSeg for Mini-dataset Spine Segmentation
• Using PaddleSeg for Lane Segmentation
• PaddleSeg in APIs
• Learn Paddleseg in 10 Mins
• Application of Interactive Segmentation Technology in Smart Mapping
• Nail art preview machine based on PaddleSeg
• Overrun monitoring of steel bar length based on PaddleSeg

For more examples, see here.

License

PaddleSeg is released under the Apache 2.0 license.

Acknowledgement

• Thanks jm12138 for contributing U²-Net.
• Thanks zjhellofss (Fu Shenshen) for contributing Attention U-Net, and Dice Loss.
• Thanks liuguoyu666, geoyee for contributing U-Net++ and U-Net3+.
• Thanks yazheng0307 (LIU Zheng) for contributing quick-start document.
• Thanks CuberrChen for contributing STDC(rethink BiSeNet), PointRend and DetailAggregateLoss.
• Thanks stuartchen1949 for contributing SegNet.
• Thanks justld (Lang Du) for contributing UPerNet, DDRNet, CCNet, ESPNetV2, DMNet, ENCNet, HRNet_W48_Contrast, FastFCN, BiSeNetV1, SECrossEntropyLoss and PixelContrastCrossEntropyLoss.
• Thanks Herman-Hu-saber (Hu Huiming) for contributing ESPNetV2.
• Thanks zhangjin12138 for contributing RandomCenterCrop.
• Thanks simuler for contributing ESPNetV1.
• Thanks ETTR123(Zhang Kai) for contributing ENet, PFPNNet.

Citation

If you find our project useful in your research, please consider citing:

@misc{liu2021paddleseg,
      title={PaddleSeg: A High-Efficient Development Toolkit for Image Segmentation},
      author={Yi Liu and Lutao Chu and Guowei Chen and Zewu Wu and Zeyu Chen and Baohua Lai and Yuying Hao},
      year={2021},
      eprint={2101.06175},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{paddleseg2019,
    title={PaddleSeg, End-to-end image segmentation kit based on PaddlePaddle},
    author={PaddlePaddle Contributors},
    howpublished = {\url{https://github.com/PaddlePaddle/PaddleSeg}},
    year={2019}
}