# Chimera: a block-based neural architecture search framework for event-based object detection

## TL;DR

Chimera is a new framework that uses neural architecture search to improve object detection with event-based cameras, achieving better performance with fewer resources.

## Contribution

Chimera introduces a block-based NAS framework tailored for event-based object detection, enabling efficient adaptation of RGB methods to event data.

## Key findings

- Chimera achieved state-of-the-art mean Average Precision on Prophesee's GEN1 dataset.
- The framework reduced model parameters by 1.6× and improved speed by 2.1×.
- It combines attention blocks, convolutions, and other architectures for effective event data processing.

## Abstract

Event-based cameras are sensors inspired by the human eye, offering advantages such as high-speed robustness and low power consumption. Established deep learning techniques have proven effective in processing event data, but there remains a significant space of possibilities that could be further explored to maximize the potential of such combinations. In this context, Chimera is a Block-Based Neural Architecture Search (NAS) framework specifically designed for Event-Based Object Detection, aiming to systematically adapt RGB-domain processing methods to the event domain. The Chimera design space is constructed from various macroblocks, including attention blocks, convolutions, State Space Models, and MLP-mixer-based architectures, providing a valuable trade-off between local and global processing capabilities, as well as varying levels of complexity. Results on Prophesee's GEN1 dataset demonstrated state-of-the-art mean Average Precision (mAP) while reducing the number of parameters by 1.6 × and achieving a 2.1 × speed-up. The project is available at: https://github.com/silvada95/Chimera.