Navigu

Visual Navigation: Users can explore images through a visually sorted 2D grid, dynamically updated as new images are retrieved based on user interactions such as dragging and zooming.

Dynamic Exploration Graph: Offers fast, high-quality approximate nearest neighbor search, outperforming other state-of-the-art methods, while ensuring connectivity and adaptability to changing datasets.

Feature Vectors: Utilizes three types of feature vectors for enhanced joint visual and textual image search, improving accuracy and efficiency. Images can be searched based on content and appearance using text queries or sample images.

Fast Visual Image Sorting: Uses the Fast Linear Assignment Sorting (FLAS) algorithm to quickly create visually sorted arrangements, ensuring a fluid and intuitive user experience.

Navigu provides a comprehensive, user-friendly solution for searching, navigating, and exploring extensive image archives with unparalleled speed and accuracy.

1. Start a Search: Enter a text query or upload a sample image to begin your visual journey. You can also click the dice to start a random search.
2. Visual Exploration: Images are displayed in a visually sorted 2D grid, grouped by similarity.
3. Interactive Map: Effortlessly explore related concepts by dragging the map in any direction.
4. Zoom In and Out: Discover more similar images or get a broader overview of the collection by zooming in and out.
5. Image Information: Simply click an image to get more information about it.
6. New Search: Double-click an image to start a new search based on that particular image .

Navigu organizes an image collection as a graph - a connection of similar images, displaying a subset at a time. Through user interactions, it is possible to navigate to related image concepts. The principle is briefly explained here:

• Similarity Graph: The image collection is organized in a hierarchical similarity graph, where each image is represented as a vertex.
• Search and Matching (a): For a visual or textual search, the best matching image is determined (shown in blue), and its edges are followed to find related images (shown in red).
• Visually Sorted Map (b): The related images are then presented in a visually sorted map, allowing for easy exploration.
• Interactive Exploration (c): Users can explore the visually sorted map by dragging and zooming.
• Dynamic Image Retrieval (d): Dragging the map may move some images out of view, leaving empty areas.
• Image Update (e): Opposite border images (shown in orange) are used to retrieve new vertices/images (shown in green) and arrange them in the empty areas.
• Seamless Integration (f): Newly retrieved images are visually arranged in the empty areas, providing a continuous and smooth exploration experience.

navigu utilizes a system of hierarchical graphs, where each higher graph level contains a quarter of the images found in the graph one level below. Zooming in or out transitions you to the next larger or smaller graph level, respectively. Dragging the view keeps you on the current graph level, simply shifting the position of interest as described before.

We currently support 5 image collections, more will follow. Currently, these are free stock photos from Pixabay, test images from the machine learning community and image collections from archives. Please contact us if you are interested in having your image collection visualized by navigu.