nvidia/LocateAnything-3B

Note

This is a remotely-sourced model from the fiftyone-locate-anything plugin, maintained by the community. It is not part of FiftyOne core and may have special installation requirements. Please review the plugin documentation and license before use.

NVIDIA LocateAnything-3B: open-vocabulary grounding VLM (3B params, Qwen2.5-3B + MoonViT + Parallel Box Decoder). Image and (frame-sampled) video. 7 operations: detect, grounding, point, scene_text, layout, text_grounding, gui_box. Returns fo.Detections or fo.Keypoints (image), {frame: label} (video)..

Details

• Model name: nvidia/LocateAnything-3B

• Model source: https://huggingface.co/nvidia/LocateAnything-3B

• Model author: NVIDIA

• Model license: NVIDIA License (non-commercial research only). See https://huggingface.co/nvidia/LocateAnything-3B

• Exposes embeddings? no

• Tags: detection, keypoints, grounding, ocr, layout, gui, torch, zero-shot, image, video, VLM, open-vocabulary

Requirements

• Packages: huggingface-hub, torch, torchvision, transformers>=4.57.1,<4.58, tokenizers>=0.22.0, Pillow, timm, lmdb, peft, opencv-python-headless>=4.10

• CPU support

  • yes

• GPU support

  • yes

Example usage

import fiftyone as fo
import fiftyone.zoo as foz

foz.register_zoo_model_source("https://github.com/Burhan-Q/fiftyone-locate-anything")

dataset = foz.load_zoo_dataset(
    "coco-2017",
    split="validation",
    dataset_name=fo.get_default_dataset_name(),
    max_samples=50,
    shuffle=True,
)

model = foz.load_zoo_model("nvidia/LocateAnything-3B")

dataset.apply_model(model, label_field="predictions")

session = fo.launch_app(dataset)