# DiffudionAnomaly

**Repository Path**: VR_NAVE/diffudionanomaly

## Basic Information

- **Project Name**: DiffudionAnomaly
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-06-23
- **Last Updated**: 2026-02-03

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## Iterative Better: Denoise Diffusion Framework for Few-shot Road Anomaly Detection

**DiffusionAnomaly is the first work of diffusion model for few-shot road anomaly detection.**

![](system.png)

Fast detection and accurate identification of road anomaly objects play a vital role in traffic safety, road maintenance, and intelligent transportation systems. Mainstream detection methods, such as Faster R-CNN and the YOLO series, are designed to achieve high-quality candidate region matching in a single inference step. Therefore, these methods are categorized as “best-oriented” approaches. However, diverse appearances, irregular shapes, blurred boundaries, and limited samples of road anomalies often cause low-quality candidate region matching, thereby hindering the performance of label assignment. To overcome these limitations, we propose a “better-oriented” framework for few-shot road anomaly detection (FSRAD), leveraging a region-refinement denoise diffusion approach. This framework reduces reliance on high-quality candidate regions, enhances generalization under data-scarce conditions, and improves adaptability to diverse anomaly shapes. We first propose an anomaly-adapted transfer pipeline to address shape sensitivity and data scarcity. In this pipeline, we design an anomaly-adapted gradient rectification strategy to improve the feature discrimination of novel anomalies while inheriting the boundary detection capability. Then, we develop a label-robust anomaly detection head, which applies a global feature enhancement and a soft matching loss optimization strategy to tackle instability from manual annotations and blurred anomaly boundaries. Experimental results demonstrate that our method achieves 10-shot novel road anomaly detection with a 6.0\% mAP improvement and a 6.4\% performance gain on the classical CODA and RDD datasets, surpassing state-of-the-art methods.


> **Iterative Better: Denoise Diffusion Framework for Few-shot Road Anomaly Detection**           
> Liangliang Cai, Qichuan Geng, Xuanqian Wang, and Zhong Zhou*

## Models
More will be released later.

### Installation

The codebases are built on top of [Detectron2](https://github.com/facebookresearch/detectron2), [Sparse R-CNN](https://github.com/PeizeSun/SparseR-CNN), and [denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch).
Thanks very much.

#### Requirements
- Linux or macOS with Python ≥ 3.6
- PyTorch ≥ 1.9.0 and [torchvision](https://github.com/pytorch/vision/) that matches the PyTorch installation.
  You can install them together at [pytorch.org](https://pytorch.org) to make sure of this
- OpenCV is optional and needed by demo and visualization

## Getting Started with DiffusionAnomaly
It will be released later.


## License

This project is under the CC-BY-NC 4.0 license. See [LICENSE](LICENSE) for details.