Skip to content
DepthWizardDepthWizard

Inference engine

predict_scene in Model_Traning/v5/infer/engine.py is the only inference path. The CLI, the FastAPI service, the hosted Space and the sliding-window evaluation all call it, so the reported metrics describe what users actually get.

Pipeline

Section titled “Pipeline”
flowchart TB
  subgraph Prepare["Prepare"]
    direction LR
    A[Scene RGB, any size] --> B[2–98 % stretch] --> C[Fill no-data] --> D[Resample to 0.5 m] --> E[512 px windows, 25 % overlap]
  end
  subgraph Predict["Predict"]
    direction LR
    F{TTA?} -- no --> G[Batched forward]
    F -- yes --> H[8 D4 views × scales] --> G
  end
  subgraph Assemble["Assemble"]
    direction LR
    I[Hann-weighted blend] --> J[Resample to input grid] --> K[height · seg · σ]
  end
  Prepare --> Predict --> Assemble

Tiling and blending

Section titled “Tiling and blending”
Scene resampled to 0.5 m512 pxstride 384 px25 %edges: reflect-padded, never zeroPer-window Hann weightwindow 1window 2window 3h(x) = Σ wᵢ(x)·hᵢ(x) / Σ wᵢ(x)seams vanish because weights fall to 0 at tile borders
Windows are 512 px with 25 % overlap (stride 384 px). Each tile's prediction is weighted by a 2-D Hann window, which falls to zero at the tile border, then normalised by the summed weights. Scene edges are reflect-padded, never zero-padded. Zero padding in v2 training crops touched about 41 % of supervised pixels.
Parameter Value
Canonical GSD 0.5 m
Window 512 × 512 px
Overlap 0.25 (stride 384 px)
Blend 2-D Hann window, normalised
Border handling reflect padding
Precision bf16 autocast on CUDA, fp32 on CPU
Batch 4 tiles by default (batch_tiles)

Test-time augmentation

Section titled “Test-time augmentation”

With TTA on, each tile runs through the 8 elements of the dihedral group D4 (4 rotations × optional flip). Each prediction is inverse-transformed and averaged, optionally at several scales (tta_scales = (1.0, 1.25)). The model checkpoint (best.pt) is chosen without TTA; TTA is applied only at evaluation and inference time.

Large scenes

Section titled “Large scenes”

Scenes above 40 megapixels go through predict_scene_windowed (run_windowed in the service). It processes horizontal bands of 2048 rows plus a one-tile margin and streams the GeoTIFF outputs straight to disk. A 20k × 20k px scene peaked at 5.9 GB RAM.

Outputs

Section titled “Outputs”
Array Units Notes
height metres above ground the nDSM, on the input grid
seg class id 0–6 argmax of Head C
std metres Head B spread; median about 0.25 m on a Cartosat-2E test scene

See Output formats for the files written to disk.

Command line

Section titled “Command line”
Model_Traning/v5/
python -m infer.predict scene.tif --ckpt outputs/v5/best.pt --tta           # nDSM (+ GeoTIFF if georeferenced)
python -m infer.predict scene.tif --ckpt outputs/v5/best.pt --absolute      # + DTM/DSM from Copernicus GLO-30
python -m infer.predict scene.tif --ckpt outputs/v5/best.pt --dem cartodem.tif
python -m infer.predict scene.png --ckpt outputs/v5/best.pt --gsd 0.5       # plain image, declared GSD
python -m infer.predict scene.tif --ckpt outputs/v5/best.pt --gcps gcps.csv # absolute via ground control points
python -m infer.predict scene.tif --ckpt outputs/v5/best.pt --report        # self-contained HTML report