Reproduce “Fine-tuning CNN Image Retrieval with No Human Annotation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Filip Radenović, Giorgos Tolias, Ondřej Chum

Researchers from the Visual Recognition Group at Czech Technical University in Prague developed a fully automated method for fine-tuning Convolutional Neural Networks for instance image retrieval, eliminating the need for human annotation. The approach, which utilizes geometric information from 3D reconstructions to generate training data, achieved new state-of-the-art performance on standard public benchmarks including Oxford Buildings and Paris.

Key insights

  • Rich geometric and camera pose information, automatically derived from 3D reconstructions, can effectively serve as a 'free' and robust source of supervision for deep learning fine-tuning.
  • Learning from 'hard' examples (challenging positive pairs with significant viewpoint changes and difficult negative pairs with visual similarities) is crucial for developing discriminative and robust image descriptors.
  • Optimizing pooling strategies through a learnable parameter (GeM) and applying discriminative post-processing (L_w) are effective architectural and algorithmic improvements for image retrieval.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.