Dataset card: structure, usage, configs + evidence

README.md

@@ -0,0 +1,308 @@
+---
+license: other
+license_name: provenance-and-license-to-follow
+task_categories:
+- depth-estimation
+- image-to-image
+tags:
+- depth
+- monocular-depth
+- rgbd
+- game-engine
+- dense-ground-truth
+- pretraining
+size_categories:
+- 10K<n<100K
+pretty_name: OriginLab Game-Depth (RGB + Dense Z-Buffer)
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*.parquet
+  - split: test
+    path: data/test-*.parquet
+  - split: extra
+    path: data/extra-*.parquet
+---
+
+# OriginLab Game-Depth: RGB + Dense Z-Buffer Depth
+
+Dense depth from game engines, as a scalable substitute for scarce real depth ground truth.
+
+## Abstract
+
+Dense depth ground truth is the bottleneck in monocular depth estimation. Real sensors are sparse, noisy,
+or indoor-only, and purpose-built synthetic datasets are expensive and narrow. Game engines already render
+a dense, exact z-buffer for every frame, for free. We ask whether that signal can stand in for real data.
+Training a depth model from scratch on ~17.8k game frames, roughly a quarter of the synthetic corpus behind
+the Lotus baseline, we find it transfers to real outdoor scenes better than that baseline (KITTI AbsRel
+0.191 vs 0.224), while indoor scenes remain its frontier. The reason is geometric rather than cosmetic: the
+game corpus teaches an outdoor ground-plane structure that real driving data shares, even though its pixels
+look nothing alike. We therefore position game z-buffers as a scalable pre-training substrate, a cheap
+geometric prior for initializing models before fine-tuning on limited real data, rather than a replacement
+for real data.
+
+*This is a preview evidence card; raw RGB is withheld pending provenance review. A few analyses are still in
+progress (the causal ablation in Section 4 and the scaling and multi-seed studies in Section 7); they will
+be added here as they complete.*
+
+## Dataset structure
+
+Each example is one RGB frame paired with dense depth from the engine z-buffer:
+
+- `image` : RGB frame (1080x1920), PNG.
+- `depth_nearness` : relative log-nearness stored as a 16-bit PNG. Decode with
+  `nearness = numpy.array(x) / 65535.0` (in [0,1], near = 1). This is relative, not metric.
+- `valid` : validity mask (0 or 255).
+- `game`, `session`, `frame`, `split` : metadata.
+
+Splits (48,615 frames total):
+
+- `train` (17,799) : the session-capped, stride-sampled curated training split used in the results below.
+- `test` (999) : session-disjoint held-out test set.
+- `extra` (29,815) : the remaining frames (these share sessions with `train`); released so you can build
+  your own curation instead of ours.
+
+### Usage
+
+```python
+from datasets import load_dataset
+import numpy as np
+
+ds = load_dataset("originlab/game-depth", split="train", streaming=True)
+ex = next(iter(ds))
+rgb      = ex["image"]                                          # PIL RGB
+nearness = np.array(ex["depth_nearness"]).astype("float32") / 65535.0   # [0,1], near = 1 (relative)
+valid    = np.array(ex["valid"]) > 0
+```
+
+## 1. Depth data is the bottleneck
+
+Every monocular-depth model is limited by the depth labels it can learn from. LiDAR is sparse and costly;
+structured-light sensors are indoor-only and noisy; pseudo-labels inherit a teacher's blind spots.
+Synthetic datasets such as Hypersim and Virtual KITTI give dense, exact depth, which is why the strongest
+open diffusion-depth models train on them, but they are hand-authored, fixed in size, and narrow in domain
+(photoreal interiors, or a driving simulator).
+
+Game engines sidestep the labeling problem entirely: the z-buffer that produces every rendered frame is
+dense per-pixel depth, available at capture time at no additional cost. Unlike a curated synthetic dataset,
+game capture is open-ended across any title, session, or environment, so the supply of dense depth grows
+with recording rather than with annotation budget. The question this card answers is whether depth learned
+from that source actually transfers to the real world.
+
+## 2. A depth dataset from game engines
+
+The dataset is 48,615 RGB frames (1080x1920) from 10 games across 98 sessions, each paired with dense
+per-pixel depth from the engine z-buffer (stored as log-nearness, `near = 1 - luma/65535`; relative, not
+metric). Splits are session-disjoint, so no scene leaks between train and test. Frames are stride-sampled to
+cut the temporal redundancy of contiguous gameplay (raw 3-fps extraction is about 21% near-duplicates; the
+sampled training split about 8%), and the training split is session-capped so no single session dominates.
+The composition is deliberately outdoor-heavy and 0% indoor, a fact that turns out to explain most of the
+results below. A per-game breakdown is in Section 8.
+
+The comparison that frames the rest of the card is with the data behind the Lotus baseline:
+
+| | Lotus training data | This dataset (ours) |
+|---|---|---|
+| Sources | Hypersim + Virtual KITTI (2 curated datasets) | 10 commercial games, 98 sessions |
+| Train size | about 74k (54k + 20k) | 48,615 total; 17,799 used here |
+| Origin | purpose-built renders / driving sim | off-the-shelf gameplay capture |
+| Scenes | indoor + road | outdoor: forest, off-road, driving, FPS, party |
+| Depth GT | metric | relative log-nearness |
+| Scaling | fixed datasets | grows with capture, not labeling |
+
+Our model is trained from scratch on roughly 4x fewer frames, entirely from games. That the resulting model
+is competitive at all is the first hint that the signal is dense and clean enough to matter.
+
+## 3. Game depth transfers to real outdoor scenes
+
+The central test is zero-shot transfer to a real benchmark the model never saw. On KITTI (1000-image
+annotated validation set), evaluated in a single fixed harness, the game-trained model has the lowest error
+of the diffusion-family models:
+
+| Model (zero-shot) | AbsRel (lower better) | delta1 (higher better) |
+|---|---|---|
+| Ours (game, from scratch) | 0.191 | 0.720 |
+| Lotus (released) | 0.224 | 0.585 |
+| Marigold | 0.244 | 0.570 |
+| Depth-Anything-V2 (real-data reference) | 0.075 | 0.947 |
+
+The gap is statistically clear, not noise: 95% bootstrap confidence intervals are [0.189, 0.194] for ours
+and [0.221, 0.226] for Lotus, which do not overlap.
+
+A model that has only ever seen rendered game frames predicts real outdoor depth more accurately than one
+trained on purpose-built synthetic data, and it does so on real photographs, which tells us synthetic-RGB
+fidelity is not the limiting factor. The advantage is not superficial: it is strongest exactly where outdoor
+scene understanding lives, on the receding ground plane and at long range, and it holds when noisy boundary
+pixels are removed, so it reflects structure the model understands rather than sensor noise it happens to
+fit. Depth-Anything-V2, a discriminative model trained on massive labeled real data, sits far ahead of all
+diffusion models; it is a reference ceiling, not a same-recipe competitor.
+
+![KITTI zero-shot, all models](figures/kitti_samples.png)
+
+Zero-shot KITTI predictions across models (inverse-depth visualization; near bright, far dark; selected
+examples). The game-trained model recovers road geometry and vehicles cleanly, ahead of the other diffusion
+models on these frames. Depth-Anything-V2 (the real-data reference) remains strongest overall (Table above).
+
+## 4. The mechanism: geometry, not appearance
+
+Why would game frames transfer to real driving scenes? The intuitive guess, that the game RGB simply looks
+like KITTI, is wrong, and measuring it is what makes the real explanation clear.
+
+Embedding every image with DINOv2 and comparing distributions, the game data is in fact closest in
+appearance to indoor NYU, not outdoor KITTI:
+
+| Pair | DINOv2 Frechet distance |
+|---|---|
+| game vs NYU | 0.98 |
+| game vs KITTI | 1.32 |
+
+If appearance drove transfer, the model would do best on NYU, the opposite of what happens. What the game
+data actually shares with KITTI is 3-D structure. Measuring the ground-plane signature of each dataset, how
+strongly distance increases from the bottom of the image to the top, the game corpus looks outdoor: its
+per-image ground-plane strength (median rho) clusters near KITTI (-0.79 vs -0.82) and far from indoor NYU
+(-0.56).
+
+![Per-image ground-plane strength](figures/geometry_profile.png)
+
+A depth model learns geometry, not texture, so the game corpus hands it an outdoor ground-plane prior that
+happens to be exactly right for real driving scenes and exactly wrong for cluttered interiors. This single
+mechanism explains the whole pattern of results: the outdoor win, the indoor gap, and the value of the data
+as a prior. To move this from a strong correlation to a causal claim, we are running the direct test:
+holding the game RGB fixed while progressively destroying the depth geometry, retraining, and watching
+transfer fall (results in progress).
+
+## 5. Indoor is the frontier
+
+The same prior that wins outdoors is a liability indoors. With no indoor frames in training, NYU is out of
+distribution, and the model loses zero-shot (AbsRel 0.149 vs Lotus 0.133). Fine-tuning on real NYU closes
+the gap. Under a matched learning-rate sweep (best checkpoint for each initialization), the game-pretrained
+model reaches AbsRel 0.116, statistically tied with the fine-tuned Lotus baseline (0.115), despite having no
+indoor data and roughly 4x less pre-training. We take the honest reading: indoor performance needs indoor
+data, and here game-pretraining matches, rather than beats, a curated-synthetic baseline. That parity is
+still a useful data-efficiency result, and it maps where the approach helps today (outdoor geometry) and
+where the next dataset version has to grow (indoor and more varied scenes).
+
+| Model | NYU AbsRel, zero-shot | NYU AbsRel, + NYU fine-tune |
+|---|---|---|
+| Ours (game, from scratch) | 0.149 | 0.116 |
+| Lotus (released) | 0.133 | 0.115 |
+| Marigold | 0.197 | not fine-tuned |
+| Depth-Anything-V2 (real-data reference) | 0.055 | not fine-tuned |
+
+(654-image Eigen test, cap 10 m, lower is better. Fine-tuned numbers use the matched learning-rate sweep,
+best checkpoint per initialization; ours and Lotus are statistically tied.)
+
+![NYU: game pre-training then fine-tuning](figures/nyu_progression.png)
+
+Game pre-training alone is weak indoors (third column, zero-shot: the outdoor prior is out of distribution
+for cluttered rooms), but it is a strong starting point. Fine-tuning on real NYU recovers the room layout
+and furniture (fourth column). Selected examples with the largest zero-shot-to-fine-tuned improvement.
+
+An earlier comparison at a single higher learning rate had suggested a larger game-pretraining advantage;
+that turned out to be an under-tuned Lotus baseline, which the matched sweep corrects. We report the
+matched, fair numbers.
+
+![NYU predictions, both models fine-tuned](figures/nyu_samples.png)
+
+NYU predictions with both models fine-tuned on real NYU under the identical recipe (selected examples).
+The game-pretrained model produces indoor depth as close to the ground truth as the fine-tuned Lotus
+baseline, consistent with the tied metrics above.
+
+## 6. What this is: a scalable pre-training substrate
+
+Read together, the results describe a specific and useful role for game-engine depth. It is not a
+replacement for real data; Depth-Anything-V2, trained on real labels, is far more accurate on the
+real-world benchmarks. It is a cheap, scalable geometric prior: dense and exact, free at capture time, and, as the KITTI result shows,
+carrying structure that transfers to the real world. The natural use is to pre-train on game depth and then
+fine-tune on whatever small real dataset a task allows, getting the benefit of a strong prior without the
+cost of collecting real dense depth.
+
+Two in-domain observations reinforce this. First, the pre-training learns genuine structure: on held-out
+game frames our model predicts depth well ahead of Lotus. Second, and more telling, the real-data model
+that dominates the benchmarks is the weakest on our frames, which means the data carries structure that
+existing models have not already absorbed.
+
+| Model on our game test set | SSI-MAE (lower better) | AbsRel |
+|---|---|---|
+| Ours (game) | 0.029 | 0.050 |
+| Lotus | 0.035 | 0.064 |
+| Marigold | 0.041 | 0.074 |
+| Depth-Anything-V2 (real-data SOTA elsewhere) | 0.055 | 0.098 |
+
+## 7. Where this goes (v0.3.0)
+
+- Scale: more frames and denser sampling, to test whether the outdoor margin widens (a data-scaling curve
+  is in progress to show whether we are still data-limited).
+- Coverage: indoor and more varied scenes, to convert the indoor frontier into a strength.
+- Confidence: multi-seed variance and confidence intervals on every number, and the causal depth-geometry
+  ablation of Section 4.
+
+## 8. Dataset composition
+
+The training split (17,799 frames) spans 10 games; no single game dominates. The held-out game test set
+(999 frames) is session-disjoint and drawn from 5 of the games.
+
+| Game | Train frames | Train % | Test frames |
+|---|---|---|---|
+| Game1 | 1,600 | 9.0 | 0 |
+| Game2 | 2,000 | 11.2 | 200 |
+| Game3 | 1,600 | 9.0 | 0 |
+| Game4 | 2,000 | 11.2 | 200 |
+| Game5 | 800 | 4.5 | 200 |
+| Game6 | 1,400 | 7.9 | 200 |
+| Game8 | 2,000 | 11.2 | 200 |
+| Game9 | 2,800 | 15.7 | 0 |
+| Game10 | 1,800 | 10.1 | 0 |
+| Game11 | 1,800 | 10.1 | 0 |
+| Total | 17,799 | 100 | 999 |
+
+Full corpus before session-capping and the train split is 48,615 frames. Machine-readable counts in
+[`results/game_distribution.json`](results/game_distribution.json).
+
+## Methodology and scope
+
+Full detail in [`METHODOLOGY.md`](METHODOLOGY.md); machine-readable metrics (including per-frame) in
+[`results/`](results/). In brief: all models run through one harness with per-model output conventions
+handled explicitly, and predictions aligned to ground truth by least-squares scale-shift. The harness is
+validated by Depth-Anything-V2 reproducing its published NYU number (about 0.055). KITTI is processed at
+native resolution, because its roughly 3.4:1 frames are otherwise squashed and blurred. The controlled
+comparison is ours vs Lotus (identical recipe, only the data differs); Marigold and Depth-Anything-V2 are
+external checkpoints included as reference points, with inference settings disclosed. Depth only; normals
+are out of scope. Point estimates are single-seed pending the variance study in Section 7.
+
+Training footprint: the trained component is the SD2-base UNet (about 0.87B trainable parameters; VAE and
+text encoder frozen), run for 6000 steps at effective batch 32, so roughly 192k images are seen, about 11
+passes over the 17,799-frame split. Compute cost is a separate axis from data amount: under a step-matched
+budget it stays fixed as the data is scaled down, so the data-utilization question (whether accuracy keeps
+rising with more data) is answered by the data-scaling curve in Section 7, not by compute; the only coupling
+is that smaller fractions imply more passes over the data (a memorization caveat for those points).
+
+## Provenance and license
+
+Provenance and license terms are being finalized and will be published here before any data release. Until
+then this is a private evidence card and raw RGB is withheld.
+
+## References
+
+1. N. Silberman, D. Hoiem, P. Kohli, R. Fergus. "Indoor Segmentation and Support Inference from RGBD
+   Images." ECCV, 2012. (NYU Depth V2)
+2. A. Geiger, P. Lenz, R. Urtasun. "Are We Ready for Autonomous Driving? The KITTI Vision Benchmark Suite."
+   CVPR, 2012. A. Geiger, P. Lenz, C. Stiller, R. Urtasun. "Vision Meets Robotics: The KITTI Dataset."
+   IJRR, 2013.
+3. M. Roberts, J. Ramapuram, A. Ranjan, et al. "Hypersim: A Photorealistic Synthetic Dataset for Holistic
+   Indoor Scene Understanding." ICCV, 2021.
+4. A. Gaidon, Q. Wang, Y. Cabon, E. Vig. "Virtual Worlds as Proxy for Multi-Object Tracking Analysis."
+   CVPR, 2016. Y. Cabon, N. Murray, M. Humenberger. "Virtual KITTI 2." arXiv:2001.10773, 2020.
+5. J. He, H. Li, W. Yin, et al. "Lotus: Diffusion-based Visual Foundation Model for High-quality Dense
+   Prediction." arXiv:2409.18124, 2024.
+6. B. Ke, A. Obukhov, S. Huang, N. Metzger, R. C. Daudt, K. Schindler. "Repurposing Diffusion-Based Image
+   Generators for Monocular Depth Estimation (Marigold)." CVPR, 2024.
+7. L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, H. Zhao. "Depth Anything V2." NeurIPS, 2024.
+   arXiv:2406.09414.
+8. M. Oquab, T. Darcet, T. Moutakanni, et al. "DINOv2: Learning Robust Visual Features without
+   Supervision." TMLR, 2023.
+
+## Citation
+
+OriginLab Game-Depth (RGB + Dense Z-Buffer), 2026. Preview version.