Limulus Vision

Overview

Limulus Vision is a research-oriented project that explores bio-inspired approaches to visual perception, drawing direct inspiration from the compound eye of the horseshoe crab (Limulus polyphemus).
Unlike human vision, the limulus visual system prioritizes motion, contrast, and temporal change over spatial resolution, making it remarkably robust in low-light and noisy environments.

This project investigates how these principles can be translated into modern computational vision systems.

Motivation

Traditional computer vision pipelines often rely on dense, uniform sampling and high-resolution imagery. While effective, these approaches can be inefficient and brittle when operating under constrained conditions.

The limulus offers an alternative model:

• Non-uniform spatial sampling
• Strong temporal integration
• Early noise suppression
• Emphasis on edges and motion gradients

Limulus Vision aims to reinterpret these biological strategies as computational primitives.

System Architecture

The system is structured as a modular perception pipeline:

1. Adaptive Sampling Layer
   Incoming frames are resampled using a non-uniform grid inspired by compound eye geometry.

2. Temporal Filtering
   Frame-to-frame differences are emphasized to enhance motion sensitivity while suppressing static noise.

3. Feature Extraction
   Edge orientation, local contrast, and motion vectors are extracted using lightweight convolutional operators.

4. Learning Layer
   PyTorch-based models are trained to classify or segment perceptual patterns using the bio-inspired representations.

Each stage is designed to be independently testable and replaceable.

Experiments & Results

Initial experiments focused on:

• Low-light motion detection
• Robust edge extraction under noise
• Comparison with standard CNN baselines

Early results show improved robustness in scenarios where conventional pipelines degrade rapidly, especially in low signal-to-noise conditions.

Future Directions

Planned next steps include:

• Extending the model to event-based vision sensors
• Integrating spiking neural network components
• Deploying the pipeline on embedded hardware
• Exploring applications in robotics and XR perception systems

This project is ongoing and actively evolving.