Skip to content

Dataset · syCen

Benchmark for evaluating video camera motion understanding

syCen/CameraBench

Image and text dataset for testing and SFT fine-tuning multimodal models to understand camera motion.

What's inside

The dataset contains between 1K and 10K examples with image and text modalities (annotated video clips are also mentioned in the card). It is created to evaluate generative and discriminative VLMs against classical SfM/SLAM methods. The authors showed that SFT fine-tuning on ~1,400 additional clips doubles the accuracy of the Qwen2.5-VL model.

What you can do with it

  • Evaluating VLM ability to recognize camera trajectories and motion
  • SFT fine-tuning of multimodal models (e.g., Qwen2.5-VL) for video analysis
  • Comparing semantic scene understanding between neural networks and classical SfM algorithms
Size
1K–10K rows
Downloads this month
9.8K
Likes
38

How to load it

from datasets import load_dataset

ds = load_dataset("syCen/CameraBench")