yangquan07 / refinenet

RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation

Multipath RefineNet

This is the source code for our paper:

RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation
Guosheng Lin, Anton Milan, Chunhua Shen, Ian Reid
In CVPR 2017

A MATLAB based framework for semantic image segmentation and general dense prediction tasks on images.

Update notes

• 23 Dec 2016: We did a major update of our code.
• (new!) 13 Feb 2018:
  1. Multi-scale prediction and evaluation code are added. We add demo files for multi-scale prediction, fusion and evaluation. Please refer to the Testing section below for more details.
  2. New models available: trained models using improved residual pooling. Available for these datasets: NYUDv2, Person_Parts, PASCAL_Context, SUNRGBD, ADE20k. These models will give better performance than the reported results in our CVPR paper.
  3. New models available: trained models using ResNet-152 for all 7 datasets. Apart from ResNet-101 based models, our ResNet-152 based models of all 7 datasets are now available for download.
  4. Updated trained model for VOC2012: this updated model is slightly better than the previous one. We previously uploaded a wrong model.
  5. All models are now available in Google Drive and Baidu Pan.
  6. More details are provided on testing, training and implementation. Please refer to Important notes in each section below.

Results

• Results on the CityScapes Dataset (single scale prediction using ResNet-101 based RefineNet)

Trained models

• (new!) Trained models for the following datasets are available for download.

1. PASCAL VOC 2012
2. Cityscapes
3. NYUDv2
4. Person_Parts
5. PASCAL_Context
6. SUNRGBD
7. ADE20k

• Downloads for the above datasets. Put the downloaded models in ./model_trained/
  • (new!) RefineNet models using ResNet-101: Google Drive or Baidu Pan
  • (new!) RefineNet models using ResNet-152: Google Drive or Baidu Pan
• Important notes:
  • The trained models of the the following datasets are using improved residual pooling: NYUDv2, Person_Parts, PASCAL_Context, SUNRGBD, ADE20k. These models will give better performance than the reported results in our CVPR paper. Please also refer to the Network architecture section below for more details about improved pooling.
  • The model for VOC2012 is updated. We previously uploaded a wrong model.

Network architecture and implementation

• You can find the network graphs that illustrate our architecture in the folder net_graphs. Please refer to our paper for more details.
• We include in this folder the details of improved residual pooling which improves the residual pooling block described in our CVPR paper.
• Important notes:
  • In our up-sampling and fusion layer, we simply use down-sampling for gradient back-propagation. Please refer to the implementation of our fusion layer for details: My_sum_layer.m.
  • please refer to our training demo files for more details on implementation

Installation

• Install MatConvNet and CuDNN. We have modified MatConvNet for our task. A modified copy of MatConvNet is provided in ./lib/. You need to compile the provided MatConvNet before running. Details of this modification and compiling can be found in main/my_matconvnet_resnet/README.md.

• An example script for exporting lib paths is main/my_matlab.sh

• Download the following ImageNet pre-trained models and place them in ./model_trained/:

  • imagenet-resnet-50-dag, imagenet-resnet-101-dag, imagenet-resnet-152-dag.

  They can be downloaded from: MatConvNet, we also have a copy in Google Drive, Baidu Pan.

Testing

1. Multi-scale prediction and evaluation (new!)

• First download the trained models and put them in ./model_trained/. Please refer to the above section Trained Models.

• Then refer to the below example scripts for prediction on your images:

  • demo_predict_mscale_[dataset name].m
  • e.g., demo_predict_mscale_voc.m, demo_predict_mscale_nyud, demo_predict_mscale_person_parts

• You may need to carefully read through the comments in these demo scripts before using.

• Important notes:

  • In the default setting, the example scripts will perform multi-scale prediction and fuse multi-scale results to generate final prediction.
  • The generated masks and scores maps will be saved in your disk. Note that the score maps are saved in the format of uint8 with values in [0 255]. You need to cast them into double and normalize into [0 1] if you want to use them.
  • The above demo files are able to perform multi-scale prediction and evaluation (e.g., in terms of IoU scores) in a single run. However, in the default setting, the performance evaluation part is disabled. Please refer to the comments in the demo files to turn on the performance evaluation.
  • Trained models using improved residual pooling will give better performance than the reported results in our CVPR paper. Please refer to the above section Trained models for more details.
  • For the images from NYUDv2 dataset, you may need to remove the white borders of the images before applying our models. More details and crop tools can be found in the NYUDv2 dataset webpage.

2. Single scale prediction and evaluation

• Single scale prediction and evaluation can be done by changing the scale setting in the multi-scale prediction demo files. Please refer the the above section for multi-scale prediction.

• We also provide simplified demo files for prediction with much less configurations. They are only for single scale prediction. Examples can be found at: demo_test_simple_voc.m and demo_test_simple_city.m.

3. Evaluation and fusion on saved results (new!)

• We provide an example script to perform multi-scale fusion on a number of predictions saved in your disk:
  • demo_fuse_saved_prediction_voc.m : fuse multiple cached predictions to generate the final prediction
• We provide an example script to evaluate the prediction masks saved in your disk:
  • demo_evaluate_saved_prediction_voc.m : evaluate the segmentation performance, e.g., in terms of IoU scores.

Training

• The following demo files are provided for training a RefineNet on your own dataset. Please carefully read through the comments in the demo files before using this training code.
  • demo_refinenet_train.m
  • demo_refinenet_train_reduce_learning_rate.m
• Important notes:
  • We use step-wise policy to reduce learning rate, and more importantly, you need to manually reduce the learning rate during the training stage. The setting of maximum training iteration just serves as a simple example and it should be adapted to your datasets. More details can be found in the comments of the training demo files.
  • We use the improved version of chained pooling in this training code, which may achieve better result than using the above provided models.

Citation

If you find the code useful, please cite our work as

@inproceedings{Lin:2017:RefineNet,
  title = {Refine{N}et: {M}ulti-Path Refinement Networks for High-Resolution Semantic Segmentation},
  shorttitle = {RefineNet: Multi-Path Refinement Networks},
  booktitle = {CVPR},
  author = {Lin, G. and Milan, A. and Shen, C. and Reid, I.},
  month = jul,
  year = {2017}
}

License

For academic usage, the code is released under the permissive BSD license. For any commercial purpose, please contact the authors.