This repository is a paper-only reproduction workflow for PocketDVDNet: Realtime Video Denoising for Real Camera Noise.

The repo mirrors the paper stages directly:

1. Prune FastDVDNet with OBProxSG
2. Retrain a Shift-Net teacher on the realistic multi-component noise model
3. Distill the final noise-map-free PocketDVDNet
4. Evaluate the distilled checkpoint

This repo uses uv only.

uv run python scripts/bootstrap.py

What bootstrap does:

• runs uv sync
• clones a pinned Shift-Net commit into external/Shift-Net
• installs that external dependency into the same environment
• checks the paper configs for unresolved paths
• warns early if required datasets or prior-stage artifacts are missing

Set DATA_ROOT before running the stages:

export DATA_ROOT=/absolute/path/to/data

Optional downloaded paper checkpoints:

./checkpoints/shift-net_retrained.pth
./checkpoints/pocketdvdnet.pt

Google Drive folder:

• https://drive.google.com/drive/folders/1Wdt6XGlgTEQQxAOoicfnk463ZP__ox1I?usp=drive_link

These checkpoints are optional convenience artifacts:

• shift-net_retrained.pth lets you skip stage 2 and go straight to stage 3.
• pocketdvdnet.pt lets you skip stage 3 and run evaluation directly.
• the paper-first path still retrains the Shift-Net teacher from scratch in stage 2.

Expected dataset layout:

$DATA_ROOT/
├── train/
│   ├── sequence_000/
│   │   ├── 00000.png
│   │   ├── 00001.png
│   │   └── ...
│   └── ...
├── DAVIS_val/
│   ├── aerial/
│   │   ├── 00000.png
│   │   ├── 00001.png
│   │   └── ...
│   └── ...
└── Set8/
    ├── snowboard/
    │   ├── 00000.png
    │   ├── 00001.png
    │   └── ...
    └── ...

Validation rules:

• training data must be clean frame sequences only
• validation sequences must contain image files directly inside each sequence directory
• all configs use repo-relative output paths and ${DATA_ROOT} for datasets
• every stage validates its required inputs before training starts

Paper-default evaluation commands:

uv run python stages/04_evaluate.py --config configs/paper/eval.yaml
uv run python stages/04_evaluate.py --config configs/paper/eval_set8.yaml
uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student.yaml
uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student_set8.yaml

These configs use the legacy paper-style seed sweep (1..99) by default. Use --fast to run the same paper protocol with the representative seed 99 only:

uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student_set8.yaml --fast

Realistic CSV-driven evaluation remains available explicitly:

uv run python stages/04_evaluate.py --config configs/paper/eval_realistic.yaml
uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student_realistic.yaml
uv run python stages/04_evaluate.py --config configs/paper/eval_set8_realistic.yaml
uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student_set8_realistic.yaml

Run the live/video demo with the published recipe and checkpoint:

uv run python scripts/live_video_inference.py --source 0
uv run python scripts/live_video_inference.py --source /path/to/video.mp4 --output-video artifacts/live/demo.mp4

The demo uses the current repo loader, a rolling 5-frame window, and tiled center-frame inference. CUDA is used automatically when available, with FP16 enabled by default on GPU for the fastest high-quality path.

Optional downloaded-checkpoint commands:

• uv run python stages/03_distill_pocketdvdnet.py --config configs/paper/distill_from_downloaded_teacher.yaml
• uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student.yaml
• uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student.yaml --fast
• uv run python stages/04_evaluate.py --config configs/paper/eval_downloaded_student_realistic.yaml
• uv run python scripts/live_video_inference.py --source 0

Optional maintainer smoke checks:

• uv run python scripts/verify_smoke.py --mode stages
• uv run python scripts/verify_smoke.py --mode checkpoints --checkpoint-dir checkpoints

Stage 1 writes:

• artifacts/prune/latest.pt
• artifacts/prune/best.pt
• artifacts/prune/pocketdvdnet_recipe.json
• artifacts/prune/summary.json

Stage 2 writes:

• artifacts/teacher/latest.pt
• artifacts/teacher/best.pt
• artifacts/teacher/summary.json

Stage 3 writes:

• artifacts/distill/latest.pt
• artifacts/distill/best.pt
• artifacts/distill/summary.json

Stage 4 writes:

• artifacts/eval/metrics.json

• PocketDVDNet.forward(x) is the final student interface. The student does not accept an explicit noise map.
• FastDVDNet is kept only as the pruning source model.
• Shift-Net is external and pinned through scripts/bootstrap.py; the old submodule flow is not used.
• stage 2 now defaults to retraining the Shift-Net teacher from scratch.
• configs/paper/pocketdvdnet_recipe.json is the committed published student recipe for optional checkpoint evaluation and resume paths.
• The recipe artifact produced in stage 1 is the contract between pruning and distillation.