Introduction

SpikeSEG is a PyTorch implementation of a biologically-inspired spiking neural network (SNN) for satellite detection and instance segmentation from event camera data. The system combines methods from four peer-reviewed publications to deliver a complete, unsupervised pipeline for neuromorphic space situational awareness.

Motivation

Traditional frame-based cameras struggle in the space domain: satellites appear as faint, sub-pixel streaks against dense star fields, and conventional sensors saturate under rapid illumination changes. Event cameras (neuromorphic sensors such as ATIS and DAVIS) solve these problems by reporting per-pixel brightness changes asynchronously with microsecond resolution, high dynamic range, and virtually no motion blur.

SpikeSEG exploits the temporal structure of these event streams with a spiking neural network that learns features without any labelled data using Spike-Timing Dependent Plasticity (STDP).

Contributions

This implementation provides:

1. A three-layer spiking convolutional encoder with Leaky Integrate-and-Fire (LIF) neurons, trained layer-wise via STDP with Winner-Take-All (WTA) competition.
2. A tied-weight decoder that maps classification spikes back to pixel space to produce saliency maps.
3. The HULK-SMASH algorithm for unsupervised instance segmentation: each classification spike is decoded individually (HULK), then instances are grouped into objects by comparing their Active Spike Hash (ASH) similarity and spatial overlap (SMASH score).
4. Volume-based evaluation following the IGARSS 2023 methodology, with informedness as the primary metric.

Pipeline Overview

flowchart LR
    A[Event Stream] --> B[Preprocessing]
    B --> C[Spiking Encoder]
    C --> D{Classification Spikes}
    D --> E[Decoder]
    E --> F[Saliency Map]
    D --> G[HULK Decoder]
    G --> H[Instance Masks]
    H --> I[SMASH Grouping]
    I --> J[Detected Objects]

Implemented Papers

#	Paper	Year	Contribution
1	Kheradpisheh et al.	2018	STDP learning rule, SNN architecture
2	Kirkland et al.	2020	SpikeSEG encoder-decoder, saliency mapping
3	Kirkland et al.	2022	HULK-SMASH instance segmentation
4	Kirkland et al.	2023	Space domain awareness, layer-wise leak (IGARSS)

Next Steps

• Installation -- set up SpikeSEG in your environment.
• Quick Start -- run a forward pass in five lines of code.
• Theoretical Background -- understand SNNs, STDP, and WTA.
• Results -- evaluation methodology and performance metrics.