Physically Consistent Trajectory-Aware Blur: a physically grounded motion-blur synthesis pipeline and a trajectory-aware dataset.

This repository contains code and materials for PC-TAB, a physically grounded motion blur synthesis framework that:

• approximates the exposure integral in linear light,
• uses three consecutive sharp frames as anchors (prev / reference / next),
• decomposes motion into camera + object residual, accounts for visibility/occlusions, rolling-shutter effects, and a lightweight ISP simulation (e.g., noise/clamp).

Built on top of this pipeline, we provide PC-TABD — a synthetic blur–sharp dataset with explicit per-pixel trajectories, along with auxiliary signals such as depth/flow and generation metadata.

The goal is to reduce the synthetic–real gap when training or fine-tuning deblurring models. Instead of simple frame averaging, we aim to reproduce key factors of real-world blur formation (exposure, trajectory shapes, rolling shutter, occlusions, ISP), while keeping the process controllable and interpretable.

PC-TABD dataset (Yandex Disk):
https://disk.yandex.ru/d/ZWUnAKowUmQSzg

This link points to the dataset download (archives / folder structure with blur–sharp pairs, trajectories, and metadata).

Folder names may slightly differ; below is the intended “mental map”.

• pc_tab/ — core synthesis library (trajectories → visibility → integration → ISP)
• configs/ — YAML/JSON configs for generation/experiments
• scripts/
  • generate_dataset.py — generate PC-TABD samples
  • augment_train.py — on-the-fly augmentation during training
  • evaluate.py — evaluation / metrics runner
• third_party/ — external dependencies / wrappers (e.g., flow/depth tools)
• docs/ — notes, figures, additional materials
• assets/ — README images
• data/ — (optional) place for dataset symlinks/caches (do not commit large files)

git clone https://github.com/illaitar/PC-TABD
cd PC-TABD

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Download from the Yandex Disk link and unpack, e.g. into ./data/pc_tabd/:

mkdir -p data/pc_tabd
# unzip / tar / etc.

python scripts/generate_dataset.py \
  --config configs/pc_tabd.yaml \
  --out data/pc_tabd_generated

TODO

This project is licensed under Creative Commons Attribution–NonCommercial 4.0 International (CC BY-NC 4.0).