# Discovering the Effectiveness of Pre-Training in a Large-scale Car-sharing Platform

**Authors:** Kyung Ho Park, Hyunhee Chung  
**arXiv:** [2305.01506](https://arxiv.org/abs/2305.01506)  
**Date:** 2023-05-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.01506v1/

## Key points

- Deep learning has greatly advanced car-sharing platforms by automating fleet management processes
- Users can now upload images of cars for inspection without physically visiting the location
- Pre-training deep neural networks is a common technique to establish effective models with limited labeled datasets
- Prior studies have not extensively explored the effectiveness of pre-training methods in dealing with the lack of labeled data for car images
- A recent study conducted analyses on various pre-training methods in image recognition tasks within a live car-sharing platform setting
- The study focused on two real-world image recognition tasks addressing many-shot and few-shot problem scenarios
- The research delved into how pre-training and fine-tuning impart distinct knowledge to neural networks for enhanced understanding
- The investigation extended into car defect recognition tasks to validate key findings and optimize performance within car-sharing platforms

## Summary

The impact of deep learning on intelligent transportation applications has greatly advanced car-sharing platforms. In the past, fleet management in car-sharing services heavily relied on human involvement. However, with modern technology, users can now upload images of cars before and after use for inspection without physically visiting the location. To automate this process, previous approaches have utilized deep neural networks with pre-training being a common technique to establish effective models with limited labeled datasets. Despite its importance in dealing with the lack of labeled data for practitioners working with car images, prior studies have not extensively explored the effectiveness of pre-training methods. In response to this gap, a recent study conducted a series of analyses to uncover the efficacy of various pre-training methods in image recognition tasks within a live car-sharing platform setting. The study focused on two real-world image recognition tasks within the platform and addressed both many-shot and few-shot problem scenarios. The analysis aimed to determine which pre-training method yielded the most effective performance under different problem settings while also delving into how pre-training and fine-tuning impart distinct knowledge to neural networks for enhanced understanding. Furthermore, the research extended its investigation into car defect recognition tasks and compared results with those from previous sections to validate key findings. Overall, this comprehensive exploration sheds light on the significance of pre-training methods in enhancing image recognition tasks within car-sharing platforms and highlights the need for further research in this area to optimize performance and efficiency.