Version history

DepthWizard was rebuilt four times. Each version responded to something the previous one measured, not to a guess.

1. v1Sep 7
   Frozen DINOv3-SAT + DPT, Head A only. A 28-minute baseline on 2×T4.
   3.32 m val
2. v2Sep 8
   Adaptive bins, segmentation, gate; SynRS3D pretraining.
   3.07 m val
   crashed at epoch 21
3. v3Sep 13
   Encoder actually unfrozen; no-padding crops, per-scene stretch, stratum balancer.
   2.61 m val (TTA)
4. v4Sep 20
   Geo calibration, ONNX, FastAPI; more data, β 0.7, soft bins.
   3.32 m GAMUS test
5. v5Sep 29
   Reverts v4's regressions; detail branch, coarse labels, forest data.
   2.48 m selection
   preliminary

What changed, and why

Version	Main change	Motivated by	Result (GAMUS val unless noted)
v1	Frozen DINOv3-SAT encoder, DPT decoder, Head A only. 11.3 M of 314.4 M params trainable	Ship a baseline first	3.321 m RMSE, r 0.856
v2	Heads B and C, gated fusion, SynRS3D → GAMUS curriculum, streaming data	The planned Phase 2 architecture	3.068 m at epoch 18. Crashed at epoch 21, and the encoder was never unfrozen
v3	Structural encoder-block lookup, full unfreeze, no-padding crops, per-scene stretch, photometric jitter, stratum balancer, flatness and normal losses	v2 post-mortem: 41 % padding, cross-sensor failure, tall bias −5.3 m	2.715 plain · 2.605 TTA · 2.723 sliding
v4	Georeferenced calibration, ONNX export, FastAPI service, landscape metrics. β 0.7, soft bins, entropy term, top-16 unfreeze, DFC23 + India data	Deliver the full product	modal: 3.631 TTA val; 3.321 GAMUS test (TTA)
DAv2	Depth Anything V2 Base backbone + pretrained DPT neck	Backbone ablation	3.432 TTA val; 3.396 GAMUS test (TTA)
v5	Revert v4’s 5 regressive settings; detail branch; coarse-label loss; corrected class ids; seeded val; DEM-anchored DSM; Cartosat input path	v4 regressed on flat ground (85.5 % of the gap); maps too smooth	2.804 seeded val (Kaggle); final run 2.481 selection score Preliminary · not test-evaluated

Best single-pass (no TTA) validation RMSE per run. Validation sets differ: v1/v2 used an earlier subset, v4 runs a harder 400-tile prefix with 2.1× more tall pixels, and v5 a seeded random 400. Treat the bars as a history, not a leaderboard; the same-protocol test comparison is on the Benchmarks page.Source: src/data/metrics.json ← Model_Traning/*/metrics.json, Research-Paper/main.tex Table III

Checkpoint lineage

v5 runs are warm-started chains, not trained from scratch. This is inferred from run directory names and resume paths:

flowchart TB
  A["v4-modal · H100 · 5 sources"] --> B["v5_probe_v4init · 2.804 m val"]
  B --> C["resume-v2 · + US3D"] --> D["resume-v3"] --> E["resume-v4-1.6 · + MVS3DM · 1.615 m MVS3DM val"]
  E --> F["v5_final_forest · H100 · + NEON · 1.545 m MVS3DM val"]
  B -. ONNX export .-> G[("depthwizard.onnx")]

Which checkpoint is where

• Published weights: DepthWizard v5 on Kaggle Models.
• ONNX graph (best.onyx/v5_probe_v4init_onnx): exported from v5_probe_v4init, not from the final run.
• v5 final run (v5_final_forest/best.pt, about 2.5 GB, stored on a Modal volume): not yet evaluated on held-out test splits and not yet exported to ONNX.
• Hosted Space: serves a DINOv3 ViT-L/16 checkpoint. Which version is live is not recorded in this repository.

Lessons that carried across versions

1. Ship a baseline first. v1 took 28 minutes and gave every later change a reference number.
2. Log metrics continuously and per stratum. Global RMSE hid v2’s padding problem and v4’s ground regression.
3. Measure before you change. Each v3 fix points to a measurement from the v2 post-mortem.
4. Change one thing at a time. v4 changed five loss and schedule settings at once, and it took a dedicated analysis (V4_modal/v3VSv4.md) to attribute the regression.
5. Don’t quote numbers you can’t reproduce. Every figure on this site comes from a metrics.json, a run.log or a results file.