Voxel51/segment_anything_video_subset51

Dataset Card for Segment Anything Video (Subset 51)

This is a FiftyOne dataset containing 917 video samples from the SA-V (Segment Anything Video) dataset. The videos are at 6 fps (matching the annotation cadence) and include both manual and automatic masklet (object mask tracklets) annotations for video object segmentation tasks.

These are the videos from Subset 51 of the full dataset.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/segment_anything_video_subset51")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

This FiftyOne dataset represents a subset of the SA-V (Segment Anything Video) dataset, designed for promptable visual segmentation (PVS) tasks. This subset contains 917 videos at 6 fps with masklet annotations - object masks tracked throughout the video - without semantic category labels. Annotations include both manually created masklets (average 3.8 per video) and automatically generated masklets from SAM 2 (average 8.9 per video).

• Curated by: Meta FAIR
• Funded by: Meta FAIR
• Shared by: harpreetsahota (FiftyOne format)
• Language(s) (NLP): Not applicable (video segmentation dataset)
• License: Creative Commons Attribution 4.0 International Public License

Dataset Sources

• Repository: https://ai.meta.com/datasets/segment-anything-video/
• Paper: https://arxiv.org/abs/2408.00714

Uses

Direct Use

This dataset is suitable for:

• Video Object Segmentation (VOS) - tracking and segmenting objects throughout video sequences
• Interactive Video Object Segmentation (iVOS) - user-guided object segmentation in videos
• Promptable Visual Segmentation (PVS) - the primary task for which SA-V was designed
• Image Segmentation - by sampling individual frames from the videos
• Training and evaluating video segmentation models - particularly models like SAM 2

The FiftyOne format enables easy exploration, visualization, filtering, and analysis of the video annotations.

Out-of-Scope Use

• Semantic segmentation with predefined categories - SA-V masklets do not include semantic category labels
• Real-time video processing at original frame rates - videos have been downsampled to 6 fps
• Audio analysis - audio has been removed during preprocessing
• Tasks requiring high temporal resolution - annotation cadence is 6 fps, not suitable for fine-grained motion analysis

Dataset Structure

FiftyOne Sample-Level Fields

Each video sample in the FiftyOne dataset contains:

• filepath - Path to the 6 fps MP4 video file
• video_id - Unique identifier (e.g., "sav_051000")
• video_environment - Environment type: "Indoor" or "Outdoor"
• video_split - Original split: "train", "val", or "test"
• video_duration - Duration in seconds
• num_manual_masklets - Count of manually annotated objects
• num_auto_masklets - Count of automatically annotated objects (0 if none)

FiftyOne Frame-Level Fields

Annotations are provided only at the 6 fps cadence (every frame in the processed videos). Each annotated frame contains up to two detection fields:

• manual - fo.Detections containing manually annotated masklets
• auto - fo.Detections containing automatically generated masklets (when available)

Detection Schema

Each fo.Detection object represents one masklet instance and contains:

• label - Always "masklet" (no semantic categories in SA-V)
• bounding_box - Normalized [x, y, w, h] tight box around the mask
• mask - Boolean instance mask array cropped to the bounding box
• masklet_id - Object ID within the video (unique per object track)
• masklet_size_bucket - Size category: "small", "medium", or "large"
• masklet_size_rel - Relative mask area (fraction of frame pixels)
• stability_score - Per-frame quality score (auto annotations only; absent in manual)

Video Format

Videos are provided at 6 fps, matching the annotation cadence. Each frame in the 6 fps videos has corresponding annotations.

Dataset Creation

Curation Rationale

This FiftyOne dataset provides the SA-V data in a structured format for exploration and experimentation with video segmentation annotations.

Source Data

Data Collection and Processing

Original Data:

• Videos provided at 6 fps matching the annotation cadence
• Annotations provided as COCO RLE format masks in JSON files

FiftyOne Processing:

• COCO RLE masks decoded to boolean arrays and cropped to bounding boxes
• Masks stored as fo.Detection objects with bounding boxes and instance masks
• Both manual and auto annotations (when available) loaded as separate detection fields

Who are the source data producers?

Crowdworkers contracted through a third-party vendor.

Annotations

Annotation process

Annotation Methods:

1. Manual Masklets - SAM 2-assisted manual annotation (average 3.8 per video)
2. Auto Masklets - Automatically generated by SAM 2 (average 8.9 per video)

Annotation Cadence:

• Annotations provided at 6 fps

Who are the annotators?

Professional annotators contracted through a third-party vendor.

Personal and Sensitive Information

• Videos subjected to face blurring
• Reports about videos can be submitted to segment-anything@meta.com

Bias, Risks, and Limitations

Technical Limitations:

• No semantic category labels (only instance masks)
• 6 fps temporal resolution
• Annotations may contain errors (both manual and automatic)
• Object selection is subjective

FiftyOne-Specific:

• This is a subset (917 samples) of the full SA-V dataset
• Mask storage format changed from COCO RLE to boolean arrays (increased storage)

Recommendations

• Consider the lack of semantic labels when designing downstream tasks
• Be mindful of storage requirements due to decoded mask format
• Use FiftyOne's filtering and visualization capabilities to explore annotations

Citation

BibTeX:

@article{ravi2024sam2,
  title={SAM 2: Segment Anything in Images and Videos},
  author={Ravi, Nikhila and Gabeur, Valentin and Hu, Yuan-Ting and Hu, Ronghang and Ryali, Chaitanya and Ma, Tengyu and Khedr, Haitham and R{\"a}dle, Roman and Rolland, Chlo{\'e} and Gustafson, Laura and Mintun, Eric and Pan, Junting and Alwala, Kalyan Vasudev and Carion, Nicolas and Wu, Chao-Yuan and Girshick, Ross and Doll{\'a}r, Piotr and Feichtenhofer, Christoph},
  journal={arXiv preprint arXiv:2408.00714},
  year={2024}
}

More Information

• SA-V dataset: https://ai.meta.com/datasets/segment-anything-video/
• FiftyOne: https://docs.voxel51.com/