# A Little Bit Attention Is All You Need for Person Re-Identification

**Authors:** Markus Eisenbach, Jannik Lübberstedt, Dustin Aganian, Horst-Michael Gross  
**arXiv:** [2302.14574](https://arxiv.org/abs/2302.14574)  
**Date:** 2023-02-28  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2302.14574v1/

## Key points

- Person re-identification is important for mobile robots to track users over extended periods, even when they are partially unobserved.
- Deep-learning based real-time feature extraction on mobile robots is often performed on special-purpose devices whose computational resources are shared for multiple tasks, making inference speed a crucial factor.
- Architectural changes that improve person re-identification often come at the cost of significantly slowing down inference, with attention blocks being one such example.
- Markus Eisenbach and his colleagues propose an attention block that only slightly affects inference speed while maintaining high accuracy levels compared to deeper networks or more complex attention blocks.
- The proposed approach involves using a novel "little bit" attention block that reduces computational complexity while maintaining high performance levels in person re-identification tasks.
- The researchers show that their approach performs well on both indoor robotic datasets and standard re-identification benchmarks.
- Real-time deep learning based feature extraction on mobile robots is challenging due to limited computational resources and power constraints. Therefore, efficient architectures are necessary for successful deployment of such systems in practical scenarios.
- The proposed "little bit" attention block offers a promising solution by reducing computational complexity while maintaining high performance levels compared to more complex architectures.
- This work has significant implications for future research in developing efficient deep learning based systems for mobile robotics applications.

## Summary

Person re-identification is a critical task in applications where mobile robots need to track users over extended periods, even when they are partially unobserved. Deep-learning based real-time feature extraction on mobile robots is often performed on special-purpose devices whose computational resources are shared for multiple tasks, making inference speed a crucial factor. However, architectural changes that improve person re-identification often come at the cost of significantly slowing down inference, with attention blocks being one such example. Some well-performing attention blocks used in the state of the art have high inference costs that make them unsuitable for mobile robotic applications. To address this issue, Markus Eisenbach and his colleagues propose an attention block that only slightly affects inference speed while maintaining high accuracy levels compared to deeper networks or more complex attention blocks. The researchers perform extensive neural architecture search to derive rules for integrating this attention block into the architecture at optimal locations to achieve the best trade-off between speed and accuracy. The proposed approach involves using a novel "little bit" attention block that reduces computational complexity while maintaining high performance levels in person re-identification tasks. The researchers show that their approach performs well on both indoor robotic datasets and standard re-identification benchmarks. The authors note that real-time deep learning based feature extraction on mobile robots is challenging due to limited computational resources and power constraints. Therefore, efficient architectures are necessary for successful deployment of such systems in practical scenarios. The proposed approach provides a promising solution by improving person re-identification without compromising on inference speed. In conclusion, the paper presents an innovative approach to address the challenge of balancing accuracy and efficiency in person re-identification tasks for mobile robotics applications. The proposed "little bit" attention block offers a promising solution by reducing computational complexity while maintaining high performance levels compared to more complex architectures. This work has significant implications for future research in developing efficient deep learning based systems for mobile robotics applications.