Benchmarking Embeddings for Similar Image Retrieval - Zucchini or Cucumber?

Explore the challenges and solutions in image-based recommendations for similar-looking food items in this conference talk from Haystack US 2024. Dive into the process of building a benchmark to evaluate image vectorization models' ability to differentiate between ambiguous items like zucchinis and cucumbers. Learn about dataset suitability for problem evaluation, selecting appropriate metrics, and communicating benchmark results effectively to team stakeholders. Discover practical tips for evaluating recommender systems in specific domains using available data, drawing from the speaker's experience in crafting a benchmark based on a hierarchical grocery store image dataset. Gain insights into the hurdles encountered during the benchmarking process and their applications to other datasets, as well as the lessons learned to improve Image Recommendation API.