Few-Shot Object Detection Using Two Stage Fine Tuning Approach with Data Augmentation

Vishwa M. Badachi, Shreya P. Inamadar, Fardeen Vaddo, Sai Satya B. V., Uday Kulkarni, Shashank Hegde

2025

Abstract

Few-Shot Object Detection (FSOD) is a specialized area within object detection that focuses on the task of identifying and localizing objects from unseen classes using only a small number of labeled examples. This is particularly important when data collection and labeling are expensive or impractical. To address this challenge, a novel two-stage fine-tuning approach combined with cutout data augmentation to improve both detection accuracy and generalization is proposed. The proposed method uses the Detectron2, a popular open-source library for object detection. The training process is divided into two stages to improve the model’s performance. To address the challenges associated with the limited availability of labeled examples, the method incorporates cutout data augmentation. Cutout augmentation involves masking random rectangular regions within training images. This augmentation technique introduces additional variation in the training data, enabling the model to focus on the important features of objects rather than overfitting to specific patterns or regions, leading to improved detection performance, especially in data-scarce scenarios. The performance of the proposed method was evaluated using the COCO dataset, a widely recognized benchmark in object detection research. Experimental results highlighted the performance gains achieved by the proposed method. Specifically, for the 10-shot setting, where only 10 labeled examples per novel class were available, the method achieved an Average Precision (AP) score of 15.7 at a high Intersection over the Union (IoU) threshold of 0.75 (AP75). The performance of the proposed methodology shows 18.8% relative improvement over the previous state-of-the-art method, demonstrating the effectiveness of the two-stage fine-tuning process combined with cutout augmentation. The proposed method tries address some of the key limitations of existing FSOD approaches.

Paper Citation

in Harvard Style

M. Badachi V., Inamadar S., Vaddo F., B. V. S., Kulkarni U. and Hegde S. (2025). Few-Shot Object Detection Using Two Stage Fine Tuning Approach with Data Augmentation. In Proceedings of the 3rd International Conference on Futuristic Technology - Volume 3: INCOFT; ISBN 978-989-758-763-4, SciTePress, pages 238-246. DOI: 10.5220/0013613000004664

in Bibtex Style

@conference{incoft25,
author={Vishwa M. Badachi and Shreya Inamadar and Fardeen Vaddo and Sai Satya B. V. and Uday Kulkarni and Shashank Hegde},
title={Few-Shot Object Detection Using Two Stage Fine Tuning Approach with Data Augmentation},
booktitle={Proceedings of the 3rd International Conference on Futuristic Technology - Volume 3: INCOFT},
year={2025},
pages={238-246},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013613000004664},
isbn={978-989-758-763-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 3rd International Conference on Futuristic Technology - Volume 3: INCOFT
TI - Few-Shot Object Detection Using Two Stage Fine Tuning Approach with Data Augmentation
SN - 978-989-758-763-4
AU - M. Badachi V.
AU - Inamadar S.
AU - Vaddo F.
AU - B. V. S.
AU - Kulkarni U.
AU - Hegde S.
PY - 2025
SP - 238
EP - 246
DO - 10.5220/0013613000004664
PB - SciTePress