A2SAM: Flaws can be Applause

[NeurIPS 2024] Flaws can be Applause: Unleashing Potential of Segmenting Ambiguous Objects in SAM

🌐 Project Homepage 📄 Paper (NeurIPS 2024) 🤗 HF Model & Dataset

Introduction

This repository contains the official implementation of our paper "Flaws can be Applause: Unleashing Potential of Segmenting Ambiguous Objects in SAM" (NeurIPS 2024).
A2SAM is a novel approach that leverages the inherent ambiguities in object segmentation to improve the performance of the Segment Anything Model (SAM).

Features

• Enhanced segmentation for ambiguous and overlapping objects
• Improved handling of complex scenes with multiple similar objects
• Built upon Meta's Segment Anything Model (SAM)
• Easy integration with existing SAM-based pipelines

Installation

git clone https://github.com/username/a2sam.git
cd a2sam
pip install -r requirements.txt

Quick Start

Training

# Train on LIDC dataset
python train_lidc.py --batch-size 16 --epochs 100 --lr 1e-4

# Train on Sim10k dataset
python train_sim10k.py --batch-size 16 --epochs 100 --lr 1e-4

Evaluation

# Evaluate on LIDC test set
python test_lidc.py --checkpoint path/to/checkpoint.pth

# Evaluate on Sim10k test set
python test_sim10k.py --checkpoint path/to/checkpoint.pth

Main Arguments

• --batch-size: Batch size for training (default: 16)
• --epochs: Number of training epochs (default: 100)
• --lr: Learning rate (default: 1e-4)
• --checkpoint: Path to model checkpoint for testing
• --data-dir: Path to dataset directory
• --output-dir: Directory to save results

For more detailed configurations, please check the argument parser in each script.

Pre-trained Models and Datasets

The pre-trained weights for the LIDC dataset and the processed Sim10k dataset are available on Hugging Face.
You can access them at the following link:

• 🤗 Hugging Face Repository

Citation

If you find this work useful in your research, please consider citing:

@inproceedings{a2sam2024,
    title={Flaws can be Applause: Unleashing Potential of Segmenting Ambiguous Objects in SAM},
    author={Chenxin Li and Yuzhi Huang and Wuyang Li and Hengyu Liu and Xinyu Liu and Qing Xu and Zhen Chen and Yue Huang and Yixuan Yuan},
    booktitle={Advances in Neural Information Processing Systems},
    year={2024},
    url={https://proceedings.neurips.cc/paper_files/paper/2024/file/50ee6db59fca8643dc625829d4a0eab9-Paper-Conference.pdf}
}