Fine-Tuning CNN Image Retrieval with No Human Annotation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are dominant in image retrieval.
• Training CNNs requires large amounts of high-quality annotated data.
• Current methods lack automated fine-tuning for large, unordered image collections.

Purpose of the Study:

• To propose a fully automated method for fine-tuning CNNs for image retrieval.
• To leverage 3D models for guided training data selection.
• To improve particular-object retrieval performance.

Main Methods:

• Fine-tuning CNNs for image retrieval using automated data selection guided by 3D models.
• Utilizing hard-positive and hard-negative examples derived from 3D model geometry and camera positions.
• Employing discriminatively learned CNN descriptor whitening.
• Introducing a novel trainable Generalized-Mean (GeM) pooling layer.

Main Results:

• Automated fine-tuning significantly enhances particular-object retrieval.
• Discriminative CNN descriptor whitening outperforms PCA whitening.
• The proposed GeM pooling layer boosts retrieval performance.
• State-of-the-art results achieved on Oxford Buildings, Paris, and Holidays datasets using the VGG network.

Conclusions:

• The proposed automated method effectively fine-tunes CNNs for image retrieval.
• Leveraging 3D model information improves retrieval accuracy.
• The novel GeM pooling layer contributes to enhanced performance.