Spaces:

TheProjectsGuy

/

AnyLoc

like 5

Runtime error

Academic project: AnyLoc: Towards Universal Visual Place Recognition

We explore interesting features learned by foundation models, specifically DINOv1 and DINOv2, that are useful for visual place recognition (image retrieval using global descriptors formed by conventional aggregation techniques) in diverse environments. We find that it is possible to learn vocabularies that generalize based on the domain of the environment (for example, a specific vocabulary that can create useful global descriptors for all aerial images).
The HuggingFace space at TheProjectsGuy/AnyLoc shows the descriptor assignment (cluster assignment) for VLAD descriptor formation. The key takeaway is that the vocabularies focus on specific items in the images that help distinguish places (for the specific domain); for example, buildings and road structures for aerial images. This application is very important for demonstrating the capabilities of our method. We also plan major application upgrades that won't be possible using the current hardware of the space.

I request a free upgrade of this space to an NVIDIA T4 GPU since the current CPU configuration (2 vCPUs) takes too long to extract DINO features (it takes over 2 minutes for two images). Interested users can try out the application here; it's currently on a local machine with a poor network (and lasts for only three days). HuggingFace community plays a crucial role in present-day academic research (in machine learning, artificial intelligence, computer vision, and many other signal processing systems); we feel supporting this application will greatly help the visual place recognition community.