SVCBench: Streaming Video Counting Benchmark

This dataset contains the clipped video segments for SVCBench, a Streaming Video Counting Benchmark for Spatial-Temporal State Maintenance. It repositions counting as a minimal, controlled probe for diagnosing how video understanding models maintain world state along the video timeline.

• Project Page: https://buaa-colalab.github.io/SVCBench/
• Code: https://github.com/buaa-colalab/SVCBench

Dataset Description

This repository provides 4,574 clipped video segments used by SVCBench, designed for evaluating spatial-temporal state maintenance capabilities in video understanding models.

Dataset Summary

• Total Videos: 4,574 clips
• Total Size: ~80 GB
• Video Format: MP4 (H.264)
• Categories: 8 subcategories across object counting and event counting tasks

Categories

Object Counting (2,297 clips):

• O1-Snap: Current-state snapshot (252 clips)
• O1-Delta: Current-state delta (98 clips)
• O2-Unique: Global unique counting (1,869 clips)
• O2-Gain: Windowed gain counting (78 clips)

Event Counting (2,277 clips):

• E1-Action: Instantaneous action (1,281 clips)
• E1-Transit: State transition (205 clips)
• E2-Periodic: Periodic action (280 clips)
• E2-Episode: Episodic segment (511 clips)

File Naming Convention

Multi-query clips

Format: {category}_{question_id}_{query_index}.mp4

Example: e1action_0000_00.mp4, e1action_0000_01.mp4

Single-query clips

Format: {category}_{question_id}.mp4

Example: o1delta_0007.mp4, o2gain_0000.mp4

Video Properties

• Encoding: H.264 (using -c copy for lossless clipping)
• Frame Rates: Preserved from source (3fps, 24fps, 25fps, 30fps, 60fps)
• Duration Accuracy: ±0.1s from annotation timestamps
• Quality: Original quality maintained (no re-encoding)

Source Datasets

Videos are clipped from multiple source datasets:

• YouTube walking tours and sports videos
• RoomTour3D (indoor navigation)
• Ego4D (first-person view)
• ScanNet, ScanNet++, ARKitScenes (3D indoor scenes)
• TOMATO, CODa, OmniWorld (temporal reasoning)
• Simulated physics videos

Usage

Loading with Python

from huggingface_hub import hf_hub_download
import cv2

# Download a specific video
video_path = hf_hub_download(
    repo_id="buaaplay/SVCBench",
    filename="e1action_0000_00.mp4",
    repo_type="dataset"
)

# Load with OpenCV
cap = cv2.VideoCapture(video_path)

Batch Download

# Install huggingface-cli
pip install huggingface_hub

# Download entire dataset
huggingface-cli download buaaplay/SVCBench --repo-type dataset --local-dir ./svcbench_videos

Annotations

For complete annotations including questions, query points, and ground truth answers, please refer to the SVCBench code repository (https://github.com/buaa-colalab/SVCBench):

• Object counting annotations: object_count_data/*.json
• Event counting annotations: event_counting_data/*.json

Each annotation file contains:

• id: Question identifier
• source_dataset: Original video source
• video_path: Original video filename
• question: Counting question
• query_time or query_points: Timestamp(s) for queries
• count: Ground truth answer(s)

Quality Validation

All videos have been validated for:

• ✓ Duration accuracy (100% within ±0.1s)
• ✓ Frame rate preservation (original fps maintained)
• ✓ No frame drops or speed changes
• ✓ Lossless clipping (no re-encoding artifacts)

Citation

If you use this dataset, please cite the SVCBench paper:

@inproceedings{liu2026svcbench,
  title     = {SVCBench: A Streaming Video Counting Benchmark for
               Spatial-Temporal State Maintenance},
  author    = {Liu, Pengyiang and Shi, Zhongyue and Hao, Hongye and Fu, Qi and
               Bi, Xueting and Zhang, Siwei and Hu, Xiaoyang and Wang, Zitian and
               Huang, Linjiang and Liu, Si},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Licensing

• Annotations (parquet/JSON) and evaluation code: CC-BY-4.0
• Self-generated physics-simulation videos: CC-BY-4.0
• Source videos: retain their respective licenses — Ego4D (Ego4D License), ScanNet / ScanNet++ (respective Terms of Use, non-commercial research), ARKitScenes (Apple ARKitScenes license), RoomTour3D (CC-BY-NC), CODa (Apache-2.0), OmniWorld (CC-BY-NC), TOMATO (academic / non-commercial only; self-created clips CC BY-NC-SA, YouTube-sourced clips under original Creative Commons licenses), and other YouTube content (platform / creator terms). The E2-Periodic clips are loop-concatenation derivatives of TOMATO and are therefore released for non-commercial research under CC BY-NC-SA, with attribution to TOMATO and the original video creators.

By downloading this dataset, you agree to adhere to CC-BY-4.0 and to the licenses of all source datasets. If you are a rights holder with any concern about specific content, please open an issue and we will address it promptly.

Dataset Statistics

Category	Clips	Avg Duration	Total Size
O1-Snap	252	~2min	~4.3 GB
O1-Delta	98	~1min	~1.7 GB
O2-Unique	1,869	~3min	~32 GB
O2-Gain	78	~1min	~1.3 GB
E1-Action	1,281	~4min	~28 GB
E1-Transit	205	~2min	~3.5 GB
E2-Periodic	280	~3min	~8.7 GB
E2-Episode	511	~2min	~4.8 GB
Total	4,574	-	~80 GB

Contact

For questions or issues, please open an issue in the SVCBench repository.