Wednesday 26 March 2025
The quest for a more efficient and accurate image-to-image (i2i) retrieval system has long been an ongoing challenge in the field of computer vision. With the proliferation of visual data, the need to quickly identify and retrieve relevant images has become increasingly important. However, traditional methods often rely on complex algorithms and extensive computational resources, making them impractical for widespread adoption.
Recently, a team of researchers made significant strides towards addressing this issue by introducing a new dataset and evaluation protocol specifically designed for i2i retrieval. ILIAS, short for Instance-Level Image Retrieval at Scale, aims to provide a standardized benchmark for measuring the performance of various models and algorithms in retrieving images based on their visual content.
The ILIAS dataset is comprised of over 100 million images across 20 categories, including products, landmarks, animals, and more. Each image is associated with multiple instance-level labels, which describe specific objects or features within the image. This unique labeling scheme enables researchers to evaluate models’ ability to retrieve images based on their visual content, rather than solely relying on textual metadata.
The evaluation protocol consists of two main tasks: image-to-image (i2i) retrieval and text-to-image (t2i) retrieval. In i2i retrieval, a model is given a query image and must return the most relevant images from the dataset. In t2i retrieval, a model is provided with a textual description of an object or feature and must retrieve the corresponding images.
The researchers evaluated various state-of-the-art models on both tasks, using the ILIAS protocol to measure their performance. The results showed that even the best-performing models struggled to achieve high accuracy, especially in certain categories. For instance, retrieval accuracy for product images was significantly lower than for landmark or animal images.
However, the dataset and evaluation protocol also revealed opportunities for improvement. By analyzing the performance of different models across various categories, researchers can identify areas where they excel and areas where they struggle. This knowledge can be used to develop more effective models and algorithms tailored to specific domains or tasks.
The introduction of ILIAS marks an important step towards advancing i2i retrieval technology. With its standardized dataset and evaluation protocol, researchers and developers now have a common framework for comparing the performance of their models and identifying areas for improvement. As the field continues to evolve, ILIAS will play a crucial role in driving innovation and pushing the boundaries of what is possible in image-to-image and text-to-image retrieval.
Cite this article: “Introducing ILIAS: A New Standard for Image-to-Image Retrieval”, The Science Archive, 2025.
Image-To-Image Retrieval, Computer Vision, Instance-Level Image Retrieval, Scale, Benchmark, Visual Content, Object Detection, Feature Extraction, Deep Learning, Text-To-Image Retrieval.
