tuhink/cambench_optical_flow

CameraBench optical flow dataset

A balanced VQA dataset for evaluating camera motion understanding in videos.

📊 Dataset Statistics

• Total Questions: 249
• Unique Videos: 70
• Unique Questions: 13
• Yes Answers: 89 (35.7%)
• No Answers: 160 (64.3%)
• Balance Ratio: 0.56
• Total Size: 258.40 MB (0.25 GB)
• Average Video Size: 3.69 MB

🎯 Task Categories

This dataset covers various camera motion tasks.

📝 Dataset Format

The dataset consists of MP4 video files with frames and optical flows stored in Parquet format.

Each record contains:

• video_name: Original video filename
• video_path: Relative path to video file (e.g., videos/video.mp4)
• frames: Sequence of extracted video frames
• optical_flows: Sequence of optical flow visualizations
• question: Binary question about camera motion
• label: Answer ("Yes" or "No")

Split: train

Statistics

• Total Questions: 5816
• Unique Videos: 207
• Unique Questions: 518
• Yes Answers: 2908 (50.0%)
• No Answers: 2908 (50.0%)
• Balance Ratio: 1.0
• Total Size: 5073.39 MB (4.95 GB)
• Average Video Size: 24.51 MB

Format: WebDataset

This split uses WebDataset format for efficient streaming:

• Tar Shards: 16 tar files
• Path: webdataset/train/train-*.tar
• Structure: Each tar contains frames, optical flows, and metadata in WebDataset format
• Usage: Load with webdataset library for streaming access

import webdataset as wds

dataset = wds.WebDataset("path/to/train-*.tar").decode("rgb")
for sample in dataset:
    video_name = sample["video_name"]
    frames = [sample[f"frame_{i:04d}.png"] for i in range(sample["num_frames"])]
    flows = [sample[f"flow_{i:04d}.png"] for i in range(sample["num_flows"])]
    # ...process sample...

Split: test

Statistics

• Total Questions: 282
• Unique Videos: 72
• Unique Questions: 12
• Yes Answers: 113 (40.1%)
• No Answers: 169 (59.9%)
• Balance Ratio: 0.6686390532544378
• Total Size: 296.45 MB (0.29 GB)
• Average Video Size: 4.12 MB

Format: Parquet

This split uses Parquet format with embedded images:

• Path: data/test-*.parquet
• Structure: Sharded parquet files with Image columns for frames and optical flows
• Usage: Load with datasets library for easy access in HuggingFace ecosystem

from datasets import load_dataset

dataset = load_dataset("your-repo-id", split="test")
for sample in dataset:
    frames = sample["frames"]  # List of PIL Images
    flows = sample["optical_flows"]  # List of PIL Images
    # ...process sample...