DVS File Type for SNN Vision Input

Overview

#

This post explains four different file formats used for storing Dynamic Vision Sensor (DVS) event data, which is essential for Spiking Neural Networks (SNN).

Four Main File Formats

#

1. Text Format (.txt)

#

Human-readable format. Each line: timestamp x y polarity

1000 120 80 1
1000 121 80 0
1001 122 81 1

Multiple events can occur at the same timestamp, enabling simultaneous event representation.

2. HDF5 Format (.h5)

#

Developed by: National Center for Supercomputing Applications (NCSA)

Widely used in scientific fields including climate modeling and astronomy.

Hierarchical Structure:

/events/
    ├── x (dataset)
    ├── y (dataset)
    ├── t (timestamp)
    └── p (polarity)
/metadata/
    ├── resolution
    └── camera_info
/analysis/
    └── statistics

Python h5py library enables efficient timestamp-based filtering.

3. AEDAT2 Format (.aedat)

#

AEDAT = Address Event Data format

Developed by the neuromorphic engineering community for event-based vision sensors.

Binary Structure: 8 bytes per event

• 4 bytes: timestamp
• 4 bytes: address (x, y, polarity encoding)

4. AEDAT4 Format (.aedat)

#

Modern packet-based format with compressed event packets for efficient streaming.

File Size Comparison

#

AEDAT4 provides the most efficient storage space.