Image Retrieval in the Wild

Type: Conference tutorial
Category: Curated resource lists / Tutorials
Event: CVPR 2020 Tutorial
Format: Slides + recorded videos (remote, via Zoom)

Overview

“Image Retrieval in the Wild” is a CVPR 2020 tutorial focused on building large-scale, practical content-based image retrieval systems for real-world applications. It emphasizes:

• Vector-based image search and approximate nearest neighbor (ANN) algorithms
• System design and deployment for billion-scale visual search
• Cross-modality person re-identification (e.g., low-res, infrared, sketch, text-to-image)
• A live implementation demo of an image search engine using a pre-trained deep model

All presentation slides and videos are accessible from the tutorial schedule.

Features

Core Topics

• Content-based image retrieval foundations

  • Role of image retrieval in interacting with large visual collections
  • Limitations of standard benchmarks (e.g., Oxford buildings) for real-world scenarios

• Approximate Nearest Neighbor (ANN) search

  • Review of state-of-the-art ANN algorithms
  • Billion-scale approximate nearest neighbor search
  • Practical guidance on selecting ANN algorithms for specific tasks and constraints

• Large-scale visual search systems

  • Case study: Online C2C marketplace app (Mercari)
  • Handling over one billion listings and ~15M monthly active users
  • System design for scalability and high availability
  • Deployment on Kubernetes for production visual search

• Heterogeneous person re-identification

  • Focus on inter-modality discrepancies as the main challenge
  • Cross-modality scenarios covered:
    • Low-resolution (LR)
    • Infrared (IR)
    • Sketch
    • Text
  • Organization and overview of available datasets per scenario
  • Summary and comparison of representative algorithms and approaches

• Live-coding demo: image search engine from scratch

  • Building a web-based image search system in ~100 lines of Python
  • Using a pre-trained deep model for feature extraction
  • End-to-end walkthrough from embedding to search interface

Access & Materials

• Tutorial held remotely (Zoom) as part of CVPR 2020
• Slides and videos linked directly from the schedule:
  • Opening & context
  • Billion-scale ANN search
  • Large-scale visual search in a C2C marketplace
  • Heterogeneous person re-identification survey
  • Live-coding demo of an image search engine

Schedule (High-Level)

• Opening & Billion-scale ANN Search – Yusuke Matsui
• Large-scale Visual Search in the Mercari C2C App – Takuma Yamaguchi
• Break
• Heterogeneous Person Re-identification Survey – Zheng Wang
• Live-coding: Image Search Engine from Scratch – Yusuke Matsui

Tags

• Tutorials
• Multimodal
• Vector search

Link

• Tutorial page: https://matsui528.github.io/cvpr2020_tutorial_retrieval/