musiq-driving-mos-v1

MUSIQ Driving MOS v1

This repository packages a custom MUSIQ-based image quality indicator fine-tuned on manually scored driving images.

What this is

• Base architecture: pyiqa MUSIQ
• Task: driving photo image quality assessment
• Training target: mos_musiq_100
• Delivery format: 5-fold grouped cross-validation ensemble

The current indicator is represented by 5 checkpoints. At inference time, run all 5 models on the same image and average their output scores to obtain the final custom MUSIQ score.

Why grouped 5-fold CV was used

The original folder split under driving_frames/train and driving_frames/val was not used as the validation protocol because the same video_id can appear in both folders, which would make validation too optimistic.

This package was validated with grouped folds:

• group_cv5_seed123_fold1
• group_cv5_seed123_fold2
• group_cv5_seed123_fold3
• group_cv5_seed123_fold4
• group_cv5_seed123_fold5

Validation Summary

Grouped 5-fold aggregate metrics:

• Mean SRCC: 0.7726
• Mean PLCC: 0.7827
• Mean KRCC: 0.6466
• Mean RMSE: 14.6473
• Mean MAE: 11.8369

Standard deviations across folds:

• SRCC std: 0.0638
• PLCC std: 0.0598
• KRCC std: 0.0597
• RMSE std: 1.5146
• MAE std: 1.2219

See eval/kfold_summary.json and eval/kfold_summary.csv for the full fold-wise breakdown.

Files

• checkpoints/fold_01_best.pth to checkpoints/fold_05_best.pth
• configs/musiq_driving.yaml
• eval/kfold_summary.json
• eval/kfold_summary.csv
• eval/fold_01_summary.json to eval/fold_05_summary.json

Score Semantics

The model was trained against mos_musiq_100, which is a MUSIQ-style score scale derived from manual ratings. The output score is therefore aligned to that transformed target rather than the original 1..5 raw labels.

Notes

• This package contains model artifacts and evaluation summaries only.
• Raw training images and manual labels are not included here by default.
• The top-level outputs/musiq_driving_cv5/best.pth from local experiments is not part of this package because it corresponds to an earlier single-fold debug run rather than the final 5-fold ensemble delivery.