Label-Efficient Semantic Segmentation with Diffusion Models

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

This code is based on datasetGAN and guided-diffusion.

Note: use –recurse-submodules when clone.

Overview

The paper investigates the representations learned by the state-of-the-art DDPMs and shows that they capture high-level semantic information valuable for downstream vision tasks.
We design a simple segmentation approach that exploits these representations and outperforms the alternatives in the few-shot operating point in the context of semantic segmentation.

Dependencies

• Python >= 3.7
• Packages: see requirements.txt

Datasets

The evaluation is performed on 6 collected datasets with a few annotated images in the training set:
Bedroom-18, FFHQ-34, Cat-15, Horse-21, CelebA-19 and ADE-Bedroom-30. The number corresponds to the number of semantic classes.

datasets.tar.gz (~47Mb)

DDPM

Pretrained DDPMs

The models trained on LSUN are adopted from guided-diffusion.
FFHQ-256 is trained by ourselves using the same model parameters as for the LSUN models.

LSUN-Bedroom: lsun_bedroom.pt
FFHQ-256: ffhq.pt
LSUN-Cat: lsun_cat.pt
LSUN-Horse: lsun_horse.pt

Run

1. Download the datasets:
     bash datasets/download_datasets.sh
2. Download the DDPM checkpoint:
      bash checkpoints/ddpm/download_checkpoint.sh <checkpoint_name>
3. Check paths in experiments/<dataset_name>/ddpm.json
4. Run: bash scripts/ddpm/train_interpreter.sh <dataset_name>

Available checkpoint names: lsun_bedroom, ffhq, lsun_cat, lsun_horse
Available dataset names: bedroom_28, ffhq_34, cat_15, horse_21, celeba_19, ade_bedroom_30

How to improve the performance

1. Set input_activations=true in experiments/<dataset_name>/ddpm.json.
      In this case, the feature dimension is 18432.
2. Tune for a particular task what diffusion steps and UNet blocks to use.

DatasetDDPM

Synthetic datasets

To download DDPM-produced synthetic datasets (50000 samples, ~7Gb):
bash synthetic-datasets/gan/download_synthetic_dataset.sh <dataset_name>

Run | Option #1

1. Download the synthetic dataset:
      bash synthetic-datasets/ddpm/download_synthetic_dataset.sh <dataset_name>
2. Check paths in experiments/<dataset_name>/datasetDDPM.json
3. Run: bash scripts/datasetDDPM/train_deeplab.sh <dataset_name>

Run | Option #2

1. Download the datasets:
      bash datasets/download_datasets.sh

2. Download the DDPM checkpoint:
      bash checkpoints/ddpm/download_checkpoint.sh <checkpoint_name>

3. Check paths in experiments/<dataset_name>/datasetDDPM.json

4. Train an interpreter on a few DDPM-produced annotated samples:
      bash scripts/datasetDDPM/train_interpreter.sh <dataset_name>

5. Generate a synthetic dataset:
      bash scripts/datasetDDPM/generate_dataset.sh <dataset_name>
       Please specify the hyperparameters in this script for the available resources.
       On 8xA100 80Gb, it takes about 12 hours to generate 10000 samples.

6. Run: bash scripts/datasetDDPM/train_deeplab.sh <dataset_name>
      One needs to specify the path to the generated data. See comments in the script.

Available checkpoint names: lsun_bedroom, ffhq, lsun_cat, lsun_horse
Available dataset names: bedroom_28, ffhq_34, cat_15, horse_21

SwAV

Pretrained SwAVs

We pretrain SwAV models using the official implementation on the LSUN and FFHQ-256 datasets:

LSUN-Bedroom: lsun_bedroom.pth
FFHQ-256: ffhq.pth
LSUN-Cat: lsun_cat.pth
LSUN-Horse: lsun_horse.pth

Training setup:

Run

1. Download the datasets:
      bash datasets/download_datasets.sh
2. Download the SwAV checkpoint:
      bash checkpoints/swav/download_checkpoint.sh <checkpoint_name>
3. Check paths in experiments/<dataset_name>/swav.json
4. Run: bash scripts/swav/train_interpreter.sh <dataset_name>

Available checkpoint names: lsun_bedroom, ffhq, lsun_cat, lsun_horse
Available dataset names: bedroom_28, ffhq_34, cat_15, horse_21, celeba_19, ade_bedroom_30

DatasetGAN

Opposed to the official implementation, more recent StyleGAN2(-ADA) models are used.

Synthetic datasets

To download GAN-produced synthetic datasets (50000 samples):

bash synthetic-datasets/gan/download_synthetic_dataset.sh <dataset_name>

Run

Since we almost fully adopt the official implementation, we don’t provide our reimplementation here.
However, one can still reproduce our results:

1. Download the synthetic dataset:
     bash synthetic-datasets/gan/download_synthetic_dataset.sh <dataset_name>
2. Change paths in experiments/<dataset_name>/datasetDDPM.json
3. Change paths and run: bash scripts/datasetDDPM/train_deeplab.sh <dataset_name>

Available dataset names: bedroom_28, ffhq_34, cat_15, horse_21

Results

• Performance in terms of mean IoU:

• Examples of segmentation masks predicted by the DDPM-based method:

Cite

@misc{baranchuk2021labelefficient,
      title={Label-Efficient Semantic Segmentation with Diffusion Models}, 
      author={Dmitry Baranchuk and Ivan Rubachev and Andrey Voynov and Valentin Khrulkov and Artem Babenko},
      year={2021},
      eprint={2112.03126},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

GitHub

View Github