Adaptive Event Camera Denoising with Spatiotemporal Density Filtering

Official implementation of the accepted ICME 2026 paper:

Adaptive Event Camera Denoising with Spatiotemporal Density Filtering

Status: Accepted by IEEE ICME 2026. The official paper version will be available after publication.

Introduction

This repository provides a training-free event camera denoising method based on spatiotemporal density filtering.

The method uses spatiotemporal Gaussian-weighted neighborhood density, adaptive thresholding, polarity-separated processing, and multi-iteration refinement to remove background activity noise and hot-pixel noise while preserving valid event structures.

Files

.
├── README.md
├── requirements.txt
├── src/
│   └── adaptive_event_denoising.py
└── scripts/
    └── visualize_event_video.py

Installation

Install the required packages:

pip install -r requirements.txt

The main dependencies are:

numpy
pandas
cupy-cuda12x
opencv-python

If you use CUDA 11, replace cupy-cuda12x with cupy-cuda11x.

Check whether CuPy can access the GPU:

python -c "import cupy as cp; print(cp.__version__); print(cp.cuda.runtime.getDeviceCount())"

Input Format

The input event file should be a CSV file without a header.

Supported formats:

x, y, p, t

or:

x, y, t

where x and y are pixel coordinates, p is event polarity, and t is timestamp.

Denoising

Run the denoising script:

python src/adaptive_event_denoising.py input.csv output.csv

Example with parameters:

python src/adaptive_event_denoising.py input.csv output.csv --L 5 --dt 1000 --base-threshold 5

If --dt is not specified, the time window will be estimated automatically.

The output is a denoised event CSV file with the same format as the input.

Visualization

An optional visualization script is provided:

python scripts/visualize_event_video.py

Before running it, modify the csv_list, labels, output_path, width, and height variables in the script according to your own data.

This script is only used to generate qualitative comparison videos and is not required for denoising.

Paper

The paper has been accepted by IEEE ICME 2026.

PDF: Coming soon.

Citation

The official citation will be updated after publication.

@inproceedings{dai2026adaptive,
  title={Adaptive Event Camera Denoising with Spatiotemporal Density Filtering},
  author={Dai, Yayu and Ge, Zhou and Li, Hengyu and Ge, Qiang},
  booktitle={IEEE International Conference on Multimedia and Expo (ICME)},
  year={2026},
  note={Accepted, to appear}
}

Contact

For questions or suggestions, please open an issue in this repository.