Awesome

<img src="figures/dinosaur.png" width="30"> Generative Semantic Segmentation

Paper

Generative Semantic Segmentation,
Jiaqi Chen, Jiachen Lu, Xiatian Zhu, and Li Zhang
CVPR 2023

Abstract

We present Generative Semantic Segmentation (GSS), a generative framework for semantic segmentation. Unlike previous methods addressing a per-pixel classification problem, we cast semantic segmentation into an image-conditioned mask generation problem. This is achieved by replacing the conventional per-pixel discriminative learning with a latent prior learning process. Specifically, we model the variational posterior distribution of latent variables given the segmentation mask. This is done by expressing the segmentation mask with a special type of image (dubbed as maskige). This posterior distribution allows to generate segmentation masks unconditionally. To implement semantic segmentation, we further introduce a conditioning network (e.g., an encoder-decoder Transformer) optimized by minimizing the divergence between the posterior distribution of maskige (i.e. segmentation masks) and the latent prior distribution of input images on the training set. Extensive experiments on standard benchmarks show that our GSS can perform competitively to prior art alternatives in the standard semantic segmentation setting, whilst achieving a new state of the art in the more challenging cross-domain setting.

TODO List

• Upload model weights and DALL-E VQVAE weight
• Provide stage-1 training code and Maskige reconstruction code
• Provide the illustration of the GSS-FF and GSS-FT-W (and more training details)
• Complete install.md
• Add dataset link

Results

Cityscapes

ADE20K

MSeg

Get Started

Environment

This implementation is build upon mmsegmentation, please follow the steps in install.md to prepare the environment.

Data

Our project is developed based on MMsegmentation. Please follow the official MMsegmentation INSTALL.md and getting_started.md for installation and dataset preparation.

Train

Since the pre-generated colors have already been provided, you can directly proceed to Latent prior learning stage.

Efficient latent posterior learning for $\mathcal{X}$ (will be released soon)

The first stage is posterior Learning, where the actual task performed is assigning a unique color to each semantic category. We propose using the Maximal distance assumption to ensure that the colors of different categories are not easily confused. To run this stage, please execute the following command:

python tools/posterior_learning.py --num_classes 150

You can use the following script to validate the color assignments for each class in your generated images. If you notice that the Intersection over Union (IoU) score for a particular class is unusually low, it may be because the assigned color for that class is too similar to the colors assigned to other classes. In such cases, you can modify the color values for that class and re-run the eval command until you are satisfied with the results. The eval command is as follows:

bash tools/dist_test.sh configs/gss/posterior_learning/dalle_reconstruction_ade20k.py ckp/non_ckp.pth 8 --eval mIoU

Latent prior learning

The pre-generated colors from latent posterior learning stage have already been provided in all configs.

# train with 8 GPUs
bash tools/dist_train.sh configs/gss/cityscapes/gss-ff_r101_768x768_80k_cityscapes.py 8
# test with 8 GPUs
bash tools/dist_test.sh configs/gss/cityscapes/gss-ff_r101_768x768_80k_cityscapes.py ./ckp_dir/iter_80000.pth 8 --eval mIoU

Reference

@inproceedings{chen2023generative,
  title={Generative Semantic Segmentation
  author={Chen, Jiaqi and Lu, Jiachen and Zhu, Xiatian and Zhang, Li},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}