segment-anything-2-hiera-small-video-torch#

Quick video segmentation model delivering rapid object tracking on standard graphics cards.

Details

  • Model name: segment-anything-2-hiera-small-video-torch

  • Model source: https://ai.meta.com/sam2/

  • Model author: Nikhila Ravi, et al.

  • Model license: Apache 2.0, BSD 3-Clause

  • Model size: 175.77 MB

  • Exposes embeddings? no

  • Tags: segment-anything, torch, zero-shot, video, transformer, official

Requirements

  • Packages: torch, torchvision, sam2

  • CPU support

    • yes

  • GPU support

    • yes

Example usage

 1import fiftyone as fo
 2import fiftyone.zoo as foz
 3from fiftyone import ViewField as F
 4
 5dataset = foz.load_zoo_dataset("quickstart-video", max_samples=2)
 6
 7# Only retain detections in the first frame
 8(
 9    dataset
10    .match_frames(F("frame_number") > 1)
11    .set_field("frames.detections", None)
12    .save()
13)
14
15model = foz.load_zoo_model("segment-anything-2-hiera-small-video-torch")
16
17# Segment inside boxes and propagate to all frames
18dataset.apply_model(
19    model,
20    label_field="segmentations",
21    prompt_field="frames.detections",  # can contain Detections or Keypoints
22)
23
24session = fo.launch_app(dataset)