NVIDIA Releases Source Code for Visual Localization Model LocateAnything-3B.

NVIDIA has released the source code for the visual localization model LocateAnything-3B.

The model can locate objects even in very dense scenes. For example, in an image with dozens of minions standing close together, it correctly highlights each one with a separate bounding box.

The main difference from most existing models is the method of generating bounding boxes. Typically, the coordinates (x1, y1, x2, y2) are predicted sequentially, which slows down the process, and errors in early stages can affect subsequent coordinates, especially when there are many objects.

LocateAnything-3B employs parallel decoding: the model predicts complete bounding boxes all at once rather than building them step by step. This makes detection more stable, particularly in scenes with a large number of objects.

The model contains 3 billion parameters and is distributed with open source code. 💜