Skip to content
Pytorch implementation for Semantic Segmentation/Scene Parsing on MIT ADE20K dataset
Python Shell
Branch: master
Clone or download
hangzhaomit Merge pull request #191 from CSAILVision/hang
Update HRNet config, model and performance
Latest commit 5c2e9f6 Aug 10, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
config update hrnet v2 Aug 10, 2019
data remove old list files Jun 4, 2019
lib fix compatibility with torch1.1 and scipy1.3 Jul 26, 2019
models add HRNet model Aug 1, 2019
teaser update teaser Apr 9, 2018
.gitignore minor re-orgnization Apr 27, 2018
LICENSE Create LICENSE Jan 16, 2019 Update Aug 10, 2019 change std Aug 3, 2019 update config Aug 3, 2019 dataset download Oct 7, 2017 replace cv2 with PIL Aug 3, 2019 minor fix Aug 3, 2019
requirements.txt add requirements Jul 5, 2019 replace cv2 with PIL Aug 3, 2019 change builder to static method Aug 1, 2019 minor fix Aug 3, 2019

Semantic Segmentation on MIT ADE20K dataset in PyTorch

This is a PyTorch implementation of semantic segmentation models on MIT ADE20K scene parsing dataset.

ADE20K is the largest open source dataset for semantic segmentation and scene parsing, released by MIT Computer Vision team. Follow the link below to find the repository for our dataset and implementations on Caffe and Torch7:

If you simply want to play with our demo, please try this link: You can upload your own photo and parse it!

All pretrained models can be found at:

[From left to right: Test Image, Ground Truth, Predicted Result]

Color encoding of semantic categories can be found here:


  • HRNet model is now supported.
  • We use configuration files to store most options which were in argument parser. The definitions of options are detailed in config/
  • We conform to Pytorch practice in data preprocessing (RGB [0, 1], substract mean, divide std).


Syncronized Batch Normalization on PyTorch

This module computes the mean and standard-deviation across all devices during training. We empirically find that a reasonable large batch size is important for segmentation. We thank Jiayuan Mao for his kind contributions, please refer to Synchronized-BatchNorm-PyTorch for details.

The implementation is easy to use as:

  • It is pure-python, no C++ extra extension libs.
  • It is completely compatible with PyTorch's implementation. Specifically, it uses unbiased variance to update the moving average, and use sqrt(max(var, eps)) instead of sqrt(var + eps).
  • It is efficient, only 20% to 30% slower than UnsyncBN.

Dynamic scales of input for training with multiple GPUs

For the task of semantic segmentation, it is good to keep aspect ratio of images during training. So we re-implement the DataParallel module, and make it support distributing data to multiple GPUs in python dict, so that each gpu can process images of different sizes. At the same time, the dataloader also operates differently.

Now the batch size of a dataloader always equals to the number of GPUs, each element will be sent to a GPU. It is also compatible with multi-processing. Note that the file index for the multi-processing dataloader is stored on the master process, which is in contradict to our goal that each worker maintains its own file list. So we use a trick that although the master process still gives dataloader an index for __getitem__ function, we just ignore such request and send a random batch dict. Also, the multiple workers forked by the dataloader all have the same seed, you will find that multiple workers will yield exactly the same data, if we use the above-mentioned trick directly. Therefore, we add one line of code which sets the defaut seed for numpy.random before activating multiple worker in dataloader.

State-of-the-Art models

  • PSPNet is scene parsing network that aggregates global representation with Pyramid Pooling Module (PPM). It is the winner model of ILSVRC'16 MIT Scene Parsing Challenge. Please refer to for details.
  • UPerNet is a model based on Feature Pyramid Network (FPN) and Pyramid Pooling Module (PPM). It doesn't need dilated convolution, an operator that is time-and-memory consuming. Without bells and whistles, it is comparable or even better compared with PSPNet, while requiring much shorter training time and less GPU memory. Please refer to for details.
  • HRNet is a recently proposed model that retains high resolution representations throughout the model, without the traditional bottleneck design. It achieves the SOTA performance on a series of pixel labeling tasks. Please refer to for details.

Supported models

We split our models into encoder and decoder, where encoders are usually modified directly from classification networks, and decoders consist of final convolutions and upsampling. We have provided some pre-configured models in the config folder.


  • MobileNetV2dilated
  • ResNet18/ResNet18dilated
  • ResNet50/ResNet50dilated
  • ResNet101/ResNet101dilated
  • HRNetV2 (W48)


  • C1 (one convolution module)
  • C1_deepsup (C1 + deep supervision trick)
  • PPM (Pyramid Pooling Module, see PSPNet paper for details.)
  • PPM_deepsup (PPM + deep supervision trick)
  • UPerNet (Pyramid Pooling + FPN head, see UperNet for details.)


IMPORTANT: The base ResNet in our repository is a customized (different from the one in torchvision). The base models will be automatically downloaded when needed.

Architecture MultiScale Testing Mean IoU Pixel Accuracy(%) Overall Score Inference Speed(fps)
MobileNetV2dilated + C1_deepsup No34.8475.7554.07 17.2
Yes33.8476.8055.32 10.3
MobileNetV2dilated + PPM_deepsup No35.7677.7756.27 14.9
Yes36.2878.2657.27 6.7
ResNet18dilated + C1_deepsup No33.8276.0554.94 13.9
Yes35.3477.4156.38 5.8
ResNet18dilated + PPM_deepsup No38.0078.6458.32 11.7
Yes38.8179.2959.05 4.2
ResNet50dilated + PPM_deepsup No41.2679.7360.50 8.3
Yes42.1480.1361.14 2.6
ResNet101dilated + PPM_deepsup No42.1980.5961.39 6.8
Yes42.5380.9161.72 2.0
UperNet50 No40.4479.8060.12 8.4
Yes41.5580.2360.89 2.9
UperNet101 No42.0080.7961.40 7.8
Yes42.6681.0161.84 2.3
HRNetV2 No42.0380.7761.40 5.8
Yes43.2081.4762.34 1.9

The training is benchmarked on a server with 8 NVIDIA Pascal Titan Xp GPUs (12GB GPU memory), the inference speed is benchmarked a single NVIDIA Pascal Titan Xp GPU, without visualization.


The code is developed under the following configurations.

  • Hardware: >=4 GPUs for training, >=1 GPU for testing (set [--gpus GPUS] accordingly)
  • Software: Ubuntu 16.04.3 LTS, CUDA>=8.0, Python>=3.5, PyTorch>=0.4.0
  • Dependencies: numpy, scipy, opencv, yacs, tqdm

Quick start: Test on an image using our trained model

  1. Here is a simple demo to do inference on a single image:
chmod +x

This script downloads a trained model (ResNet50dilated + PPM_deepsup) and a test image, runs the test script, and saves predicted segmentation (.png) to the working directory.

  1. To test on an image or a folder of images ($PATH_IMG), you can simply do the following:
python3 -u --imgs $PATH_IMG --gpu $GPU --cfg $CFG


  1. Download the ADE20K scene parsing dataset:
chmod +x
  1. Train a model by selecting the GPUs ($GPUS) and configuration file ($CFG) to use. During training, checkpoints by default are saved in folder ckpt.
python3 --gpus $GPUS --cfg $CFG 
  • To choose which gpus to use, you can either do --gpus 0-7, or --gpus 0,2,4,6.

For example, you can start with our provided configurations:

  • Train MobileNetV2dilated + C1_deepsup
python3 --gpus GPUS --cfg config/ade20k-mobilenetv2dilated-c1_deepsup.yaml
  • Train ResNet50dilated + PPM_deepsup
python3 --gpus GPUS --cfg config/ade20k-resnet50dilated-ppm_deepsup.yaml
  • Train UPerNet101
python3 --gpus GPUS --cfg config/ade20k-resnet101-upernet.yaml
  1. You can also override options in commandline, for example python3 TRAIN.num_epoch 10 .


  1. Evaluate a trained model on the validation set. Add VAL.visualize True in argument to output visualizations as shown in teaser.

For example:

  • Evaluate MobileNetV2dilated + C1_deepsup
python3 --gpus GPUS --cfg config/ade20k-mobilenetv2dilated-c1_deepsup.yaml
  • Evaluate ResNet50dilated + PPM_deepsup
python3 --gpus GPUS --cfg config/ade20k-resnet50dilated-ppm_deepsup.yaml
  • Evaluate UPerNet101
python3 --gpus GPUS --cfg config/ade20k-resnet101-upernet.yaml


If you find the code or pre-trained models useful, please cite the following papers:

Semantic Understanding of Scenes through ADE20K Dataset. B. Zhou, H. Zhao, X. Puig, T. Xiao, S. Fidler, A. Barriuso and A. Torralba. International Journal on Computer Vision (IJCV), 2018. (

  title={Semantic understanding of scenes through the ade20k dataset},
  author={Zhou, Bolei and Zhao, Hang and Puig, Xavier and Xiao, Tete and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio},
  journal={International Journal on Computer Vision},

Scene Parsing through ADE20K Dataset. B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso and A. Torralba. Computer Vision and Pattern Recognition (CVPR), 2017. (

    title={Scene Parsing through ADE20K Dataset},
    author={Zhou, Bolei and Zhao, Hang and Puig, Xavier and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
You can’t perform that action at this time.