Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021, official Pytorch implementation)

Paper

Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen.

Abstract

Self-training is a competitive approach in domain adaptive segmentation, which trains the network with the pseudo labels on the target domain. However inevitably, the pseudo labels are noisy and the target features are dispersed due to the discrepancy between source and target domains. In this paper, we rely on representative prototypes, the feature centroids of classes, to address the two issues for unsupervised domain adaptation. In particular, we take one step further and exploit the feature distances from prototypes that provide richer information than mere prototypes. Specifically, we use it to estimate the likelihood of pseudo labels to facilitate online correction in the course of training. Meanwhile, we align the prototypical assignments based on relative feature distances for two different views of the same target, producing a more compact target feature space. Moreover, we find that distilling the already learned knowledge to a self-supervised pretrained model further boosts the performance. Our method shows tremendous performance advantage over state-of-the-art methods.

Installation

Install dependencies:

pip install -r requirements.txt

Data Preparation

Download Cityscapes, GTA5 and SYNTHIA-RAND-CITYSCAPES.

Inference Using Pretrained Model

1) GTA5 -> Cityscapes

Download the pretrained model (57.5 mIoU) and save it in ./pretrained/gta2citylabv2_stage3. Then run the command

python test.py --bn_clr --student_init simclr --resume ./pretrained/gta2citylabv2_stage3/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl

2) SYNTHIA -> Cityscapes

Download the pretrained model (55.5 mIoU, 62.0 mIoU for 16, 13 categories respectively) and save it in ./pretrained/syn2citylabv2_stage3. Then run the command

python test.py --bn_clr --student_init simclr --n_class 16 --resume ./pretrained/syn2citylabv2_stage3/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl

Training

To reproduce the performance, you need 4 GPUs with no less than 16G memory.

1) GTA5 -> Cityscapes

Stage1. Download warm-up model (43.3 mIoU), and save it in ./pretrained/gta2citylabv2_warmup/.

Generate soft pseudo label.

python generate_pseudo_label.py --name gta2citylabv2_warmup_soft --soft --resume_path ./pretrained/gta2citylabv2_warmup/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast

Calculate initial prototypes.

python calc_prototype.py --resume_path ./pretrained/gta2citylabv2_warmup/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl

Train stage1.

python train.py --name gta2citylabv2_stage1Denoise --used_save_pseudo --ema --proto_rectify --moving_prototype --path_soft Pseudo/gta2citylabv2_warmup_soft --resume_path ./pretrained/gta2citylabv2_warmup/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --proto_consistW 10 --rce --regular_w 0.1

Stage2. This stage needs well-trained model from stage1 as teacher model. You can get it by above command or download the pretrained model stage1 model(53.7 mIoU) and save it in ./pretrained/gta2citylabv2_stage1Denoise/ (path of resume_path). Besides, download the pretrained model simclr model and save it to ./pretrained/simclr/.

Generate pseudo label.

python generate_pseudo_label.py --name gta2citylabv2_stage1Denoise --flip --resume_path ./logs/gta2citylabv2_stage1Denoise/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast

Train stage2.

python train.py --name gta2citylabv2_stage2 --stage stage2 --used_save_pseudo --path_LP Pseudo/gta2citylabv2_stage1Denoise --resume_path ./logs/gta2citylabv2_stage1Denoise/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --S_pseudo 1 --threshold 0.95 --distillation 1 --finetune --lr 6e-4 --student_init simclr --bn_clr --no_resume

Stage3. This stage needs well-trained model from stage2 as the teacher model. You can get it with the above command or download the pretrained model stage2 model(56.9 mIoU) and save it in ./pretrained/gta2citylabv2_stage2/ (path of resume_path).

Generate pseudo label.

python generate_pseudo_label.py --name gta2citylabv2_stage2 --flip --resume_path ./logs/gta2citylabv2_stage2/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast --bn_clr --student_init simclr

Train stage3.

python train.py --name gta2citylabv2_stage3 --stage stage3 --used_save_pseudo --path_LP Pseudo/gta2citylabv2_stage2 --resume_path ./logs/gta2citylabv2_stage2/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --S_pseudo 1 --threshold 0.95 --distillation 1 --finetune --lr 6e-4 --student_init simclr --bn_clr --ema_bn

2) SYNTHIA -> Cityscapes

Stage1. Download warmup model(41.4 mIoU), save it in ./pretrained/syn2citylabv2_warmup/.

Generate soft pseudo label.

python generate_pseudo_label.py --name syn2citylabv2_warmup_soft --soft --n_class 16 --resume_path ./pretrained/syn2citylabv2_warmup/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast

Calculate initial prototypes.

python calc_prototype.py --resume_path ./pretrained/syn2citylabv2_warmup/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --n_class 16

Train stage1.

python train.py --name syn2citylabv2_stage1Denoise --src_dataset synthia --n_class 16 --src_rootpath src_rootpath --used_save_pseudo --path_soft Pseudo/syn2citylabv2_warmup_soft --ema --proto_rectify --moving_prototype --proto_consistW 10 --resume_path ./pretrained/syn2citylabv2_warmup/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --rce

Stage2. This stage needs well-trained model from stage1 as teacher model. You can get it by above command or download released pretrained stage1 model(51.9 mIoU) and save it in ./pretrained/syn2citylabv2_stage1Denoise/ (path of resume_path).

Generate pseudo label.

python generate_pseudo_label.py --name syn2citylabv2_stage1Denoise --flip --resume_path ./logs/syn2citylabv2_stage2/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast --n_class 16

Train stage2.

python train.py --name syn2citylabv2_stage2 --stage stage2 --src_dataset synthia --n_class 16 --src_rootpath src_rootpath --used_save_pseudo --path_LP Pseudo/syn2citylabv2_stage1Denoise --resume_path ./logs/syn2citylabv2_stage2/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --S_pseudo 1 --threshold 0.95 --distillation 1 --finetune --lr 6e-4 --student_init simclr --bn_clr --no_resume

Stage3. This stage needs well-trained model from stage2 as teacher model. You can get it by above command or download released pretrained stage2 model(54.6 mIoU) and save it in ./pretrained/stn2citylabv2_stage2/ (path of resume_path).

Generate pseudo label.

python generate_pseudo_label.py --name syn2citylabv2_stage2 --flip --resume_path ./logs/syn2citylabv2_stage2/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast --bn_clr --student_init simclr --n_class 16

Train stage3.

python train.py --name syn2citylabv2_stage3 --stage stage3 --src_dataset synthia --n_class 16 --src_rootpath src_rootpath --used_save_pseudo --path_LP Pseudo/syn2citylabv2_stage2 --resume_path ./logs/syn2citylabv2_stage2/from_synthia_to_cityscapes_on_deeplabv2_best_model.pkl --S_pseudo 1 --threshold 0.95 --distillation 1 --finetune --lr 6e-4 --student_init simclr --bn_clr --ema_bn

Citation

If you like our work and use the code or models for your research, please cite our work as follows.

@article{zhang2021prototypical,
    title={Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation},
    author={Zhang, Pan and Zhang, Bo and Zhang, Ting and Chen, Dong and Wang, Yong and Wen, Fang},
    journal={arXiv preprint arXiv:2101.10979},
    year={2021}
}

License

The codes and the pretrained model in this repository are under the MIT license as specified by the LICENSE file.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Acknowledgments

This code is heavily borrowed from CAG_UDA.
We also thank Jiayuan Mao for his Synchronized Batch Normalization code.

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021)

Related tags

Overview

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021, official Pytorch implementation)

Paper

Abstract

Installation

Data Preparation

Inference Using Pretrained Model

Training

Citation

License

Acknowledgments

Owner

Microsoft

Deep Learning agent of Starcraft2, similar to AlphaStar of DeepMind except size of network.

A lightweight library to compare different PyTorch implementations of the same network architecture.

Speech Separation Using an Asynchronous Fully Recurrent Convolutional Neural Network

Official implementation for "QS-Attn: Query-Selected Attention for Contrastive Learning in I2I Translation" (CVPR 2022)

Python KNN model: Predicting a probability of getting a work visa. Tableau: Non-immigrant visas over the years.

Use MATLAB to simulate the signal and extract features. Use PyTorch to build and train deep network to do spectrum sensing.

Elastic weight consolidation technique for incremental learning.

Fully Convolutional Refined Auto Encoding Generative Adversarial Networks for 3D Multi Object Scenes

Real-Time Social Distance Monitoring tool using Computer Vision

Wav2Vec for speech recognition, classification, and audio classification

The Pytorch implementation for "Video-Text Pre-training with Learned Regions"

Implementation of Neural Distance Embeddings for Biological Sequences (NeuroSEED) in PyTorch

Pytorch implementation of Distributed Proximal Policy Optimization: https://arxiv.org/abs/1707.02286

Numenta Platform for Intelligent Computing is an implementation of Hierarchical Temporal Memory (HTM), a theory of intelligence based strictly on the neuroscience of the neocortex.

code for paper -- "Seamless Satellite-image Synthesis"

Official implementation of the paper Label-Efficient Semantic Segmentation with Diffusion Models

Flybirds - BDD-driven natural language automated testing framework, present by Trip Flight

FSL-Mate: A collection of resources for few-shot learning (FSL).

Code implementing "Improving Deep Learning Interpretability by Saliency Guided Training"

Adapter-BERT: Parameter-Efficient Transfer Learning for NLP.