Jelly3000/OmniSTVG

BOSTVG Dataset

The BOSTVG Dataset is designed for research on Spatio-Temporal Omni-Object Video Grounding (OmniSTVG). It contains video clips and high-quality spatio-temporal annotations for grounding and tracking tasks.

1. Dataset Overview

1.1 Video Files

Videos are grouped into three semantic categories and distributed as three compressed archives:

File	Category	# Videos	Description
videos_animal.zip	Animal	4,138	Animal-related videos
videos_human.zip	Human	3,615	Human activity videos
videos_machine.zip	Machine	2,288	Machines and equipment videos

After extraction, the directory structure is: {category}/{subcategory}/{video_name}.mp4. Examples:

animal/shrimp/BREEDING_Shrimp_How_Many_in_30_Days-seg5.mp4
human/sumo/Big_Show_attempts_to_overpower_sumo_champion_Akebono_at-seg1.mp4
machine/car/Driving_Downtown_-_New_York_City_4K_-_USA-seg20.mp4

• Top-level category: animal, human, or machine
• Subcategory: fine-grained object or activity category (e.g., shrimp, sumo, car)
• Video key (JSON key): subcategory/video_name.mp4

1.2 Annotation Files

File	Split	# Videos
annotations/train.json	Training Set	8,106
annotations/test.json	Test Set	1,912

2. Annotation Format

The annotation file is a JSON dictionary where:

• Key: video path (subcategory/video_name.mp4)
• Value: annotation object corresponding to the video clip

2.1 Video-Level Fields

Field	Type	Description
st_time	float	Start timestamp of the clip (seconds)
ed_time	float	End timestamp of the clip (seconds)
st_frame	int	Start frame index
end_frame	int	End frame index
fps	int / float	Frame rate
img_num	int	Total number of frames in the video
width	int	Video width (pixels)
height	int	Video height (pixels)
caption	str	Natural language description of the video clip
targets	list[str]	List of target object names (one-to-one correspondence with tracks)
tracks	list[object]	Multi-object frame-level trajectories

2.2 Track Format

Each element in tracks has the following structure:

{
  "id": 1,
  "object": "sumo wrestler",
  "box": {
    "762": [188.0, 410.0, 406.0, 876.0],
    "763": [197.89, 414.54, 406.32, 870.83]
  }
}

Field	Description
id	Unique object identifier
object	Object category name
box	Mapping from frame index (string) to bounding box coordinates

2.3 Bounding Box Coordinates

Bounding boxes are represented as: [x1, y1, x2, y2], where:

• Coordinates are normalized to the range [0, 1000]
• Values should be scaled using the corresponding video width and height
• Missing annotations are represented as:[null, null, null, null]
• Frame indices range from st_frame to end_frame (inclusive)

2.4 Example Annotation

{
  "sumo/Big_Show_attempts_to_overpower_sumo_champion_Akebono_at-seg1.mp4": {
    "st_time": 26.28,
    "ed_time": 31.34,
    "st_frame": 762,
    "end_frame": 909,
    "fps": 29,
    "img_num": 947,
    "width": 854,
    "height": 480,
    "caption": "Two sumo wrestlers grappling, pushing and trying to overpower each other.",
    "targets": ["sumo wrestler", "sumo wrestler"],
    "tracks": [
      {
        "id": 1,
        "object": "sumo wrestler",
        "box": {
          "762": [188.0, 410.0, 406.0, 876.0],
          "763": [197.89, 414.54, 406.32, 870.83]
        }
      }
    ]
  }
}

3. Usage Notes

1. Extract all three videos_*.zip archives into the same root directory while preserving the structure: {category}/{subcategory}/{video}.mp4.
2. Construct video paths using the JSON key. For example:

key: sumo/xxx-seg1.mp4
path: ../human/sumo/xxx-seg1.mp4

3. To access the bounding box of a specific frame: tracks[i]["box"][str(frame_id)].
4. Convert normalized coordinates to pixel coordinates:

x_pixel = x_norm * width / 1000
y_pixel = y_norm * height / 1000

4. Dataset Statistics

Item	Count
Training Videos	8,106
Test Videos	1,912
Total Videos	10,018