| # Methodology and Reproducibility |
|
|
| ## Datasets |
| - Game-depth (ours): 48,615 RGB (1080x1920) plus dense engine z-buffer depth (log-nearness). Training uses |
| a session-balanced split (at most 200 frames/session, 17,799 frames), stride-sampled to reduce 3-fps |
| temporal redundancy. Splits are session-disjoint. Held-out game test set: 999 session-disjoint frames. |
| - NYU Depth V2 (real, indoor): official BTS split, 24,231 train frames |
| (`nyudepthv2_train_files_with_gt.txt`), 654 Eigen test. Depth = uint16 PNG / 1000 m; valid = png > 0. |
| - KITTI (real, outdoor): depth-selection `val_selection_cropped`, 1000 frames, 352x1216, annotated |
| semi-dense GT. Depth = uint16 PNG / 256 m; valid = png > 0; eval cap 80 m. |
|
|
| ## Model and training recipe |
| - Architecture and recipe: Lotus (latent diffusion, SD2-base backbone; single-step x0 at t=999; RGB+depth |
| latent concat, 8-channel conv_in; `trunc_disparity` normalization). Recipe used verbatim; only the data |
| reader differs. |
| - Game pre-training (our model): init SD2-base; effective batch 32; 6000 steps; LR 3e-5 constant. |
| - NYU fine-tuning: identical recipe for both initializations (our game model, released Lotus). Both are |
| evaluated under a learning-rate sweep (best-of checkpoints saved every 750 steps). A first comparison at |
| LR 3e-5 suggested a game-pretraining advantage, but a fair LR sweep showed the Lotus baseline had been |
| under-tuned (best LR-1e-5 checkpoint reaches AbsRel 0.115). The honest conclusion is that the indoor |
| fine-tuning difference is a training-recipe effect, not a data effect; the matched sweep for our own model |
| is being finalized. |
|
|
| ## Evaluation harness (all models, one protocol) |
| - Inference: LotusGPipeline for the diffusion models, single-step, task-emb depth. |
| - Per-model output conventions handled explicitly: ours and Lotus predict nearness (a disparity proxy), |
| Marigold predicts affine-invariant depth, Depth-Anything-V2 predicts inverse depth. |
| - Per-model inference settings, disclosed for fairness: ours and Lotus single-step; Marigold 10 steps with |
| ensemble size 1 (single-sample, not its multi-sample best); Depth-Anything-V2 a single forward pass. |
| - Resolution: NYU and game frames processed at long-side 768. KITTI processed at native resolution, because |
| its roughly 3.4:1 frames are otherwise squashed to about 224 px tall and blurred; native processing |
| improves every model and is the pinned KITTI protocol. |
| - Metrics: AbsRel, SqRel, RMSE, RMSElog, log10, delta1/2/3, plus SSI-MAE and boundary-F1 for the game |
| held-out set. |
| - Cropping and caps: NYU Eigen crop, cap 10 m. KITTI no extra crop (val set is pre-cropped), cap 80 m. |
| - Alignment: least-squares scale-shift to the ground truth (predictions are relative). |
| - Harness validation: Depth-Anything-V2 reproduces its published NYU AbsRel (about 0.055) in this harness, |
| so the relative ordering across models is trustworthy even though single-step numbers differ from each |
| paper's own protocol. |
|
|
| ## Analysis |
| - Appearance (DINOv2 Frechet distance): frozen DINOv2 ViT-S/14 embeddings, L2-normalized; Frechet distance |
| between Gaussian-fit embedding sets (game training split vs NYU vs KITTI). Preprocessing is |
| aspect-preserving center-crop to 224 (a naive square resize was tested and rejected because it distorts |
| wide KITTI). This is a semantic, not photometric, similarity measure and is treated as one correlational |
| signal, not proof. |
| - Depth geometry (ground-plane strength): for each image, Spearman rank correlation rho between pixel row |
| and pixel distance over valid pixels (subsampled to 4000 px/image); we report the per-image rho |
| distribution (median and mean). Distance is depth (m) for real sets and 1 minus nearness for the game |
| data. Being rank-based and per-image, rho is scale-invariant and unaffected by KITTI's sky-crop. |
| - Causal test (in progress): keep the game RGB fixed, progressively flatten the depth target to destroy the |
| ground-plane geometry, retrain from scratch at each corruption level, and measure KITTI transfer. If |
| transfer degrades as geometry is destroyed, the geometry is the causal driver rather than a confound. |
|
|
| ## KITTI win robustness |
| The KITTI advantage is checked against the concern that a metric win could reflect fitting sensor noise or a |
| trivial ground plane. Recomputing the win under stricter valid masks refutes this: it grows when boundary |
| and isolated LiDAR pixels are eroded (+0.074 AbsRel gap) and at long range beyond 20 m (+0.086), and is |
| neutral only on upper-image vertical structures. That is the signature of a genuine ground-plane advantage. |
|
|
| ## Files |
| - `results/results_summary.json`: headline metrics. |
| - `results/geometry_analysis.json` and `results/geo_rho_per_image.json`: geometry rho. |
| - `results/fid_recheck.json` and `results/domain_analysis.json`: DINOv2 Frechet distance. |
| - `figures/geometry_profile.png`: per-image ground-plane rho distribution. |
|
|
| ## Known caveats |
| - Indoor is out of distribution (0% indoor in training); the NYU zero-shot gap reflects data distribution. |
| - Depth GT is relative log-nearness, not metric; cross-dataset comparisons are done in normalized space. |
| - Point estimates are single-seed; multi-seed variance and bootstrap confidence intervals are being added. |
| - Marigold and Depth-Anything-V2 are external checkpoints trained under conditions we do not control, so |
| they are reference points rather than controlled comparisons; the controlled comparison is ours vs Lotus. |
|
|
