AntoineGuedon/Surflo-eval-data

Surflo evaluation data

Preprocessed Tanks and Temples scenes for evaluating Surflo, a consistent 3D surface flow model with global state.

Surflo's evaluation consumes preprocessed caches, not raw images: the frozen VGGT-1B backbone is run once per scene offline and its tokens are stored, so scripts/evaluate.py can score a checkpoint without re-running the backbone. This repository is that cache.

Contents

Six T&T scenes: Barn, Caterpillar, Courthouse, Ignatius, Meetingroom, Truck, cached at four input-view counts.

View count	Path	Samples / scene	Size
4	tnt/4views/	1	1.7 GB
8	tnt/8views/	1	2.0 GB
16	tnt/16views/	2	3.9 GB
32	tnt/32views/	1	3.9 GB
	total		13 GB

Each view count is a self-contained directory; you only need to download the one you intend to evaluate at.

Layout

tnt/<N>views/<scene>/
 ├── sample_0000_views_<NNN>.pt    # one cache per (sample, view count)
 └── surface_data.npz              # ground-truth surface cloud

What a sample_*.pt holds

Key	Shape	Notes
aggregated_tokens_list	(4, N, 745, 2048) at layers 4 / 11 / 17 / 23	VGGT-1B intr layers are None
patch_start_idx	5	where patch tokens begin
rgb_images	(1, N, 3, 280, 518)	input views, fp16, [0, 1]
vggt_world_points	(1, N, 280, 518, 3)	back-projected VGGT pointmap, fp16
vggt_extrinsics / vggt_intrinsics	(1, N, 3, 4) / (1, N, 3, 3)	VGGT-frame ca
colmap_extrinsics / colmap_intrinsics	(N, 3, 4) / (N, 3, 3)	COLMAP-frame ca
alignment_L / alignment_T	(1, 3, 3) / (1, 3)	COLMAP - VGGT affine: p_vggt = p_colmap @ L + T
scene_center / scene_radius	(1, 3) / scalar	scene extent
image_names	list[N]	source filenames

surface_data.npz stores the ground-truth surface as 200 chunks of 50,000 oriented points, i.e. 10 M points and normals per scene (n_chunks, then points_NNN / normals_NNN), shuffled so any prefix is an unbiased sample. These stay in the COLMAP frame; the evaluation aligns predictions to them with Umeyama + robust ICP.

Download

pip install -U huggingface_hub

# One view count
hf download AntoineGuedon/Surflo-eval-data --repo-type dataset \
    --include "tnt/16views/*" --local-dir ./Surflo-eval-data

# Everything (13 GB)
hf download AntoineGuedon/Surflo-eval-data --repo-type dataset \
    --local-dir ./Surflo-eval-data

# A single scene, to try it out
hf download AntoineGuedon/Surflo-eval-data --repo-type dataset \
    --include "tnt/16views/Ignatius/*" --local-dir ./Surflo-eval-data

Downloads resume if interrupted.

Usage

From a Surflo checkout, point data_dir at the view-count directory:

# Plain (unguided) flow
python scripts/evaluate.py benchmarks=tnt mode=plain \
    ckpt=/path/to/surflo_v0.pt \
    data_dir=./Surflo-eval-data/tnt/16views \
    output_json=eval_results/tnt_plain.json

# With rendering guidance
python scripts/evaluate.py benchmarks=tnt mode=guided \
    ckpt=/path/to/surflo_v0.pt \
    data_dir=./Surflo-eval-data/tnt/16views \
    num_query_points=200000 \
    output_json=eval_results/tnt_guided.json

The checkpoint lives at AntoineGuedon/Surflo-v0.

The same caches also score the feed-forward baselines through the identical alignment and metric core:

python scripts/evaluate.py benchmarks=tnt predictor=vggt \
    data_dir=./Surflo-eval-data/tnt/16views \
    output_json=eval_results/tnt_vggt.json

Provenance and licensing

This is a derivative of the Tanks and Temples benchmark. It contains downsampled source imagery (518x280), camera parameters, ground-truth surface samples derived from the benchmark's reference reconstructions, and features computed by VGGT-1B.

Component	Source	License
Source imagery, cameras, GT surfaces	[Tanks and Temples](https://www.tanksandtemples.
aggregated_tokens_list features	VGGT-1B	CC BY-NC 4.0

Released under CC BY-NC 4.0. The non-commercial term is inherited: the cached features are outputs of VGGT-1B, whose weights are CC BY-NC 4.0, and Surflo itself is released under the non-commercial Gaussian-Splatting License. No ownership is claimed over the underlying Tanks and Temples captures, which remain the property of their authors under their own terms.

If you use this data, cite Tanks and Temples as well as Surflo.

Citation

@article{knapitsch2017tanks,
  title   = {Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction},
  author  = {Knapitsch, Arno and Park, Jaesik and Zhou, Qian-Yi and Koltun, Vladlen},
  journal = {ACM Transactions on Graphics},
  volume  = {36},
  number  = {4},
  year    = {2017}
}

@article{guedon2026surflo,
  title   = {Surflo: Consistent 3D Surface Flow Model with Global State},
  author  = {Gu{\'e}don, Antoine and Nakamura, Shu and Dufour, Nicolas
             and Lei, Jiahui and Nishino, Ko and Kanazawa, Angjoo},
  journal = {arXiv preprint arXiv:2606.13644},
  year    = {2026}
}