LForster/Sintel-Low-light-Noise

Sintel Low-light Noise ELD

A synthetic low-light optical flow dataset derived from MPI Sintel using the ELD low-light noise preset.

This dataset contains noisy RGB frames for both the train and test splits and is intended for:

• optical flow robustness evaluation in low-light conditions
• fine-tuning pretrained optical flow models
• controlled experiments on synthetic low-light degradation

Contents

The dataset contains ELD-corrupted Sintel frames for:

• training
• test

This dataset includes only the noisy ELD data.

What This Dataset Is For

This dataset is useful when you want to test or train optical flow models on darker, noisier Sintel-style inputs without changing the underlying scene content.

Typical use cases:

• compare model performance on standard vs low-light inputs
• fine-tune a pretrained model for low-light robustness
• benchmark robustness under synthetic low-light degradation

Noise Model

This dataset uses the ELD low-light noise model.

The ELD corruption includes:

• brightness reduction
• shot noise
• read noise
• quantization noise
• banding artifacts

Compared with more aggressive synthetic corruption models, ELD generally produces more stable and visually plausible low-light results.

Why Sintel + ELD?

MPI Sintel is widely used for optical flow evaluation, but it does not natively include low-light variants.

Applying ELD-style degradation provides:

• a controlled robustness benchmark
• the same scene/layout content as Sintel
• a direct way to study low-light failure modes in optical flow

Recommended Use

Best use:

• fine-tune a pretrained optical flow model
• evaluate robustness to low-light corruption
• compare against clean Sintel performance

Less recommended:

• treating this as real-world low-light ground truth
• relying on it as the only low-light training source

This dataset is synthetic and is best used for controlled experiments.

File Structure

Sintel-noisy/
  train/
    alley_1/
    alley_2/
    ...
  test/
    ambush_1/
    cave_3/
    ...

Notes

• This dataset contains only the ELD noisy version.
• It is a synthetic low-light corruption dataset, not a real capture dataset.
• Transfer to real low-light video should be validated separately.

Acknowledgements

This dataset is derived from MPI Sintel and applies synthetic ELD low-light corruption to the original image content.

Please respect the licensing terms of the original Sintel dataset.