How Much Does Input Curation Alone Improve 3DGS from Casual Video?

Code release for the SIGGRAPH Asia 2026 poster "How Much Does Input Curation Alone Improve 3D Gaussian Splatting from Casual Video?" (Oshio et al., SA Posters '26, Kuala Lumpur).

Paper (3-page abstract, PDF) · Project page · DOI: 10.1145/3829333.3847887

This repository contains (1) the evaluated frame-extraction tool and (2) the complete A/B benchmark protocol used in the paper, so that every number in the abstract can be reproduced end to end.

[video] -> extract_frames.py -> [images/ + colmap_run.sh] -> COLMAP -> 3DGS

Contents

Path	Purpose
extract_frames.py	The evaluated selector: parallax-triggered, sharpness-filtered keyframe extraction with self-configured COLMAP scripts. Single file, OpenCV + NumPy only. See README_tool.md.
benchmark/prepare_scene.py	Video -> held-out test views (every 2 s, sharpest in ±5) + training sets for both arms (Ours / Uniform at the same frame budget, same resolution cap).
benchmark/run_benchmark.py	One-scene runner: COLMAP (identical settings for both arms, automatic fallback), gsplat simple_trainer (15 k iters, default hyperparameters, fixed seed), metrics. Resumable.
benchmark/metrics.py	PSNR / SSIM / LPIPS on the held-out views (same torchmetrics configs as the trainer).
benchmark/make_tables.py	Aggregation, common-test-view pairing, cluster-bootstrap 95% CIs, LaTeX table.
benchmark/make_blur.py	GoPro/REDS-style synthetic exposure blur (linear-space temporal averaging of 60 fps video; all levels re-encoded identically).
benchmark/run_blur_batch.sh	Dose-response experiment driver (blur-0/2/4/8, shared sharp test views).
benchmark/extra_metrics.py	Coverage-stratification analysis (per-view nearest-training-view distance), SfM statistics, Wilcoxon tests.
benchmark/run_seed_check.py	Seed-robustness spot check (seeds 42/43/44, both arms).
benchmark/extra_perceptual.py	DISTS / FLIP over the eval renders.
benchmark/gsplat_examples/	Vendored gsplat v1.5.3 examples with minimal patches (see PATCHES.md; identical for both arms).
benchmark/results/	Aggregated outputs backing the paper: summary.csv, extra_metrics.json, seed_check.json, perceptual.json.

Reproducing

1. Environment. Windows or Linux host with an NVIDIA GPU (paper: RTX 3080, 10 GB). Install COLMAP (>= 3.12; note that vocabulary trees must be the faiss format, e.g. vocab_tree_faiss_flickr100K_words256K.bin from the COLMAP releases). Python 3.10+ with benchmark/requirements.txt; the training venv is created by benchmark/setup_wsl_env.sh (torch 2.4.1+cu124, gsplat 1.5.3 prebuilt wheel, torchmetrics).
2. Data. Place each scene as data/scenes/<name>/input.mp4. The paper uses the Tanks and Temples source videos (Barn, Horse, Ignatius), five DL3DV-Benchmark captures (hashes in benchmark/paper_scenes.txt; original videos from DL3DV/DL3DV-ALL-video), and two casual phone captures.
3. Run. python benchmark/run_all.py (scenes are processed independently and resumably), then python benchmark/make_tables.py and python benchmark/extra_metrics.py.
4. Blur study. bash benchmark/run_blur_batch.sh cpu then ... gpu.

Fairness rules implemented by the protocol: identical frame budget per scene, identical COLMAP configuration for both arms (chosen by the tool's own QC), joint registration of held-out test frames, metrics computed only on test views registered in both arms, fixed seed, and no blur handling in the uniform arm (the standard ffmpeg recipe).

License: MIT (see LICENSE).

Citation

@inproceedings{oshio2026curation,
  author    = {Oshio, Yuki and Fukiharu, Yuya and Shiki, Akane and
               Kiyoyama, Yukito and Kihara, Miki and Ito, Ai and Imai, Shota},
  title     = {How Much Does Input Curation Alone Improve 3D Gaussian
               Splatting from Casual Video?},
  booktitle = {SIGGRAPH Asia 2026 Posters (SA Posters '26)},
  year      = {2026},
  publisher = {ACM}
}