Skip to content

Latest commit

 

History

History
105 lines (78 loc) · 3.32 KB

README.md

File metadata and controls

105 lines (78 loc) · 3.32 KB

CVAE-SM

This repository contains the code for the paper "CVAE-SM: A Conditional Variational Autoencoder with Style Modulation for Efficient Uncertainty Quantification." The work focuses on enhancing uncertainty quantification in object segmentation, particularly within challenging underwater environments.

Framework Illustration of the proposed framework for uncertainty quantification in underwater object segmentation. The framework is comprised of four primary modules: the Prior Encoder and Posterior Encoder generate conditional latent codes, the Style Modulator adjusts latent codes and content features, and the Decoder generates final segmentation maps. During testing, we sample latent codes from a normal distribution guided by the mean and standard deviation of the Prior encoder. This approach yields a diverse array of segmentation maps, facilitating robust uncertainty quantification.

Prerequisites

  • Python 3.9
  • PyTorch 2.4.1+cu121
  • Torchvision
  • Albumentations
  • OpenCV
  • Matplotlib
  • tqdm

This method was tested in:

  • GPU: Tesla V100 with 32 GB memory

Usage

Clone the repository:

git clone https://github.com/aminullah6264/CVAE-SE.git
cd CVAE-SM

Installation

To create a conda environment with the required packages, use the following command:

conda env create -f cvae.yml

Dataset

You can download the dataset used for training and testing from Google Drive. Make sure to extract the dataset and place it in the Dataset directory before running the training script.

Dataset/
├── train/
│   ├── images/
│   └── masks/
├── test/
│   ├── images/
│   ├── whole/
│   │   └── masks/
│   ├── severity_lv1/
│   ├── severity_lv2/
│   ├── severity_lv3/
│   └── severity_lv4/

Dataset Example images from the dataset, demonstrating different levels of perturbations applied to underwater images for uncertainty quantification.

Training

To train the model with all default values, use the following command:

python train.py --DATASET_PATH ./Dataset/train \
                --SAVE_DIR ./Weights \
                --NUM_EPOCHS 120 \
                --NUM_CLASSES 5 \
                --LEARNING_RATE 1e-4 \
                --BATCH_SIZE 6 \
                --NUM_ENSEMBLES 2

Testing

python test.py --dataset_dir ./Dataset/test_val/ \
               --model_checkpoint ./Weights/trained_model_20240926_214052/CVAE_SE_FCN_with_Modulation_SingleGPU.pth \
               --results_dir ./Results/ \
               --num_ensembles 10

Visualization

The visualization code will be uploaded shortly. Stay tuned for updates!

Citation

If you find this code useful, please cite our paper:

@article{ullah2024cvae_sm,
  author={Ullah, Amin and Yan, Taiqing and Fuxin, Li},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={CVAE-SM: A Conditional Variational Autoencoder with Style Modulation for Efficient Uncertainty Quantification}, 
  year={2024},
  volume={},
  number={},
  pages={10786-10792},
  doi={10.1109/ICRA57147.2024.10611160}
  }