CameraBench Video Dataset For Camera Motion Analysis

CameraBench is a dataset for understanding camera motions in videos, created by syCen. It includes approximately 1,400 annotated video clips used for fine-tuning models like Qwen2.5-VL. The dataset supports evaluation of Structure-from-Motion, Visual Language Models, and scene-aware semantic analysis.

Use Cases

• Fine-tune generative VLMs like Qwen2.5-VL on 1,400 annotated clips to improve camera motion classification.
• Benchmark Structure-from-Motion methods against VLMs using VQAScore for pure geometry tasks.
• Evaluate discriminative VLMs using CLIPScore/ITMScore on camera motion recognition.
• Analyze scene-aware semantic cues in video that are missed by classical SfM/SLAM techniques.

Strengths

• Contains approximately 1,400 annotated video clips for supervised fine-tuning.
• Enables direct comparison between classical SfM/SLAM and modern VLMs on camera motion tasks.
• Supports evaluation with multiple metrics including VQAScore, CLIPScore, and ITMScore.

Limitations

• Dataset size is relatively small with only about 1,400 clips, limiting large-scale model training.
• Specific column structure, file formats, and total video count are not publicly detailed.
• Potential bias towards the types of camera motions and scenes present in the annotated subset.

Provenance

Source

syCen via Hugging Face

Collection Method

Collection and annotation of video clips for camera motion analysis.

Time Range

null

Freshness

Last updated on 2025-07-30.

Geography

null