LianeMarilin/4k-video-annotations

4K Video Annotations — Shot Segmentation and Camera Motion

This dataset contains 12 frame-accurate shot clips segmented from five short cinematic video sequences. Every clip is paired with a detailed, manually reviewed annotation covering visible content, subject actions, shot scale, camera angle, camera movement, movement direction, stabilization, composition, lighting, color, pacing, transitions, timecodes, and technical properties.

The footage depicts a tense nighttime sequence on and around a rain-soaked glass roof, featuring human figures and a large dinosaur-like creature. Annotation field names are in English; descriptive annotation values are in Chinese.

Dataset at a Glance

Property	Value
Source sequences	5
Segmented shots	12
Annotation records	12
Fields per record	34
Total frames	771
Total duration	32.157126 seconds
Resolution	3840 × 2160
Frame rate	24000/1001 fps (approximately 23.976)
Video encoding	H.264, YUV 4:2:0
Color/HDR	BT.2020 + PQ, HDR mastering metadata retained
Audio	No audio stream in the source or segmented clips
Review status	Frame-by-frame human review

Visual Gallery

scene-0604 shot 002 High-angle overhead confrontation.	scene-0605 shot 001 Short rain-soaked action insert.
scene-0606 shot 002 Fast directional subject movement.	scene-0607 shot 001 Single continuous cinematic shot.
scene-0608 shot 002 The longest shot in the dataset.	scene-0608 shot 004 Late-sequence action and reframing.

Each gallery image is a contact sheet sampled from one segmented clip. The repository also includes a contact sheet for every shot under 镜头预览/.

Shot Distribution

Source sequence	Shot count	Frame ranges (end exclusive)
scene-0604.mp4	3	[0, 24), [24, 50), [50, 75)
scene-0605.mp4	1	[0, 15)
scene-0606.mp4	2	[0, 6), [6, 18)
scene-0607.mp4	1	[0, 68)
scene-0608.mp4	5	[0, 104), [104, 357), [357, 458), [458, 572), [572, 595)

Segmentation Method

A shot is defined as a continuous sequence between two adjacent hard cuts. Boundaries use the first frame of the new shot as the next clip's start, and end timecodes are exclusive.

All candidate transitions were reviewed frame by frame. Lightning creates strong exposure changes in this footage; these flashes were not marked as shot boundaries when composition and motion remained continuous.

The segmented clips were frame-accurately re-encoded while retaining:

• 3840 × 2160 resolution;
• 24000/1001 frame rate;
• BT.2020 color primaries;
• SMPTE ST 2084 (PQ) transfer characteristics;
• HDR mastering-display and content-light metadata.

Annotation Schema

Each line of 镜头标注.jsonl is one UTF-8 JSON object describing one continuous shot.

Group	Fields
Identity and paths	clip_id, source_video, clip_path, shot_index
Frame and time boundaries	start_frame, end_frame_inclusive, frame_count, start_timecode, end_timecode_exclusive, duration_seconds
Technical properties	resolution, fps, audio
Editing	transition_in, transition_out
Scene context	scene, time_of_day, weather
Visible subjects	subjects, subject_actions
Content description	content_summary, detailed_description
Cinematography	shot_size, camera_angle, camera_movement, movement_direction, stabilization, composition
Visual style	lighting, color_tone, visual_pacing
Additional review	on_screen_text, quality_notes, review_status

Important Field Semantics

• start_frame and end_frame_inclusive are zero-based source-frame indices.
• end_timecode_exclusive marks the first instant outside the current shot.
• subjects is an array of visible generic subject categories.
• camera_movement records fixed, pan, tilt, push, pull, tracking, lateral movement, orbiting, or compound movement when visually supported.
• movement_direction describes the observed camera or compositional direction.
• stabilization separates stable, mostly stable, handheld-feeling, and dynamically unstable shots.
• review_status records the manual frame-level review state.

No character names, performer identities, or source-film identities are inferred.

Example Record

{
  "clip_id": "scene-0604_shot-001",
  "source_video": "scene-0604.mp4",
  "clip_path": "镜头片段/scene-0604/scene-0604_shot-001.mp4",
  "shot_index": 1,
  "start_frame": 0,
  "end_frame_inclusive": 23,
  "frame_count": 24,
  "start_timecode": "00:00:00.000",
  "end_timecode_exclusive": "00:00:01.001",
  "duration_seconds": 1.001,
  "resolution": "3840×2160",
  "fps": "24000/1001 (23.976)",
  "audio": "无音轨",
  "transition_in": "源视频起点",
  "transition_out": "硬切",
  "content_summary": "男子在湿滑的玻璃屋顶上躲避逼近的恐龙。",
  "shot_size": "中全景",
  "camera_angle": "略低机位、侧向斜角",
  "camera_movement": "快速横向跟摇，伴随轻微手持震动",
  "movement_direction": "主要向右跟随主体动作",
  "stabilization": "动态不稳定",
  "review_status": "已逐帧复核"
}

The displayed example is shortened for readability. The actual JSONL record contains all 34 required fields.

Repository Structure

.
├── README.md
├── 镜头标注.jsonl              # 12 shot-level annotation records
├── 标注字段契约.json            # Machine-readable 34-field contract
├── 镜头标注说明.md              # Annotation guidance in Chinese
├── 镜头片段/
│   ├── scene-0604/             # 3 MP4 clips
│   ├── scene-0605/             # 1 MP4 clip
│   ├── scene-0606/             # 2 MP4 clips
│   ├── scene-0607/             # 1 MP4 clip
│   └── scene-0608/             # 5 MP4 clips
├── 镜头预览/                    # 12 contact sheets + one overview
├── assets/                     # README gallery images
├── 交付清单.json
├── 要求矩阵.md
├── 质量检查报告.json
└── 质量检查报告.md

There is no Excel file in this dataset.

Loading the Annotations

from datasets import load_dataset

dataset = load_dataset(
    "LianeMarilin/4k-video-annotations",
    data_files="镜头标注.jsonl",
    split="train",
)

print(dataset[0]["clip_id"])
print(dataset[0]["camera_movement"])

To download the video referenced by an annotation record:

from huggingface_hub import hf_hub_download

video_path = hf_hub_download(
    repo_id="LianeMarilin/4k-video-annotations",
    repo_type="dataset",
    filename=dataset[0]["clip_path"],
)

print(video_path)

Quality Assurance

• 12/12 annotation records parse successfully.
• Every record contains exactly 34 fields with no missing or empty values.
• 12/12 relative clip paths resolve to non-empty MP4 files.
• The segmented clips account for 771/771 source frames with no gap or overlap.
• 12/12 clips pass full-frame decoding.
• 12/12 clips retain the expected resolution, frame rate, color metadata, and HDR metadata.
• The five original source files were not modified during segmentation.

Detailed machine-readable and human-readable QA reports are included in the repository.

Intended Uses

This compact dataset can support:

• shot-boundary detection experiments;
• camera-movement and cinematography classification;
• video-language retrieval and grounding prototypes;
• shot-level captioning and visual-description evaluation;
• frame-accurate video-processing pipeline tests;
• HDR metadata preservation checks.

Because the dataset is small and visually narrow, it should be treated as a curated evaluation or demonstration set rather than a representative training corpus.

Limitations and Responsible Use

• The footage covers one cinematic scenario and is not visually diverse.
• The dataset has no audio.
• Descriptive annotations are human-authored in Chinese and may reflect reasonable visual judgment.
• Strong lightning flashes and dark HDR imagery may challenge automatic boundary detectors.
• Rights and licensing for the underlying footage are not declared by this dataset card. Users are responsible for verifying that their intended use is permitted and for respecting applicable copyright, privacy, and platform rules.

Version

Initial public release: 2026-09-14.