Dates are commit dates from the repository history (2026).
| Date |
Milestone |
| Sep 30 |
Telecom coverage and flood response scenarios in the web app |
| Sep 29 |
v5 final run on a Modal H100: forest data (NEON), forest and sparse checkpoint selection, best selection score 2.481 m |
| Sep 29 |
New React frontend merged; repository restructured |
| Sep 27 |
ONNX export of v5_probe_v4init with a parity check; v5 validation report gallery |
| Sep 26 |
Measured GSDs: GAMUS 0.25 m (not 0.33 m), US3D 0.32 m; US3D added to training; warm-started v5 chain |
| Sep 25 |
v5 Kaggle probe runs; New Delhi scenes split out of DFC23 as the India set |
| Sep 24 |
v5 implementation: detail branch, coarse-label loss, uncertainty output, unfreeze ramp, label audit, reference validation in the viewer; same-protocol v3/v4 evaluation |
| Sep 22 |
Depth Anything V2 backbone ablation (DAV2_V1) |
| Sep 20 |
v4-modal run on H100 completes (GAMUS test 3.32 m with TTA) |
| Sep 18 |
H100 / Modal optimisation |
| Sep 14–16 |
v4 on Kaggle 2 × T4 with DDP; DDP unused-parameter fix in the deepest fusion block |
| Sep 13 |
v3 H100 run: 2.715 m plain, 2.605 m TTA on GAMUS validation |
| Sep 11 |
v4 code: geo calibration, viewer, FastAPI service, domain adaptation |
| Sep 11 |
v3 fine-tuning code |
| Sep 8 |
v2 training framework (adaptive bins, segmentation, gate) |
| Sep 7 |
v1 baseline notebook (frozen encoder, 3.32 m) |
| Date |
Milestone |
| Aug 31 |
Repository created; literature review and early notebooks |
