Semi-Supervised Auto-Labeling and Generative Data Augmentation Pipeline Using Modified OverFeat Architectures for Multi-Camera 3D Lane and Object Perception Datasets

• Abey city

  Lautech

  Author

The development of robust perception systems for autonomous driving is fundamentally constrained by the high cost and labor-intensive nature of annotating 3D lane and object detection datasets. While deep learning architectures have demonstrated remarkable capabilities in vehicle and lane detection, their performance remains contingent upon access to large-scale, accurately labeled datasets that capture diverse driving scenarios . Manual annotation of multi-camera 3D data is prohibitively expensive, with lane instance labeling averaging 90 minutes per image for fine annotations . This research presents a novel semi-supervised auto-labeling and generative data augmentation pipeline built upon a modified OverFeat CNN architecture to address the data scarcity challenge. The proposed framework integrates an automatic label generation mechanism leveraging HD map priors, a semi-supervised learning paradigm that combines limited labeled data with abundant unlabeled samples, and a generative data augmentation module employing diffusion-based synthesis to produce diverse driving scenes . Experimental validation on a comprehensive multi-camera dataset demonstrates that models trained with our pipeline achieve 89.4% of fully supervised performance using only 15% of manual labels, while the generative augmentation component improves rare-class detection by 34.7% . The framework operates at real-time inference speeds exceeding 10 Hz on standard GPU configurations, making it suitable for practical deployment. This research contributes a replicable, label-efficient methodology that significantly reduces annotation costs while maintaining high perception accuracy, with implications for accelerating autonomous driving system development across diverse operational domains .

Copyright (c) 2026 Abey city (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.