Berlin model — capture-holdout experiment 20 (93% usable on a photograph it never saw)

The Berlin model from capture-holdout experiment 20 (commit 0c59820), the best model so far on the held-out-photograph ruler.

mission score	0.078 (0 = a usable fix on every frame, 1 = silence, 2 = confidently wrong everywhere)
usable fixes	93.0% of held-out frames: confident and within 100 m
confidently wrong	0.75%
abstains	6.25%
median miss	36 m
scored on	Berlin TrueDOP 2024, a photograph the model never saw in training (it trained on Brandenburg DOP20 and TrueDOP 2022/2023)
size	31.7 MB fp32 ONNX (about 8 MB once converted to int8, not yet done)
host latency	48.8 ms

Interface. Input frame: float32 [1, 3, 128, 128], RGB in 0–1, a straight-down view covering about 128 m of ground (about 100 m altitude with a typical lens). Output uvc: [1, 3] = normalised map position (u, v) and confidence. Below confidence 0.3 the model abstains. Conversion to latitude/longitude for the Berlin box (UTM 33N origin 385883, 5821316; 6,939 × 6,828 m) is in the "How to use it" section of the project overview.

What it is not. It knows one bounding box of Berlin and nothing else. On a held-out block of Berlin it never trained on, it gives 0.5% usable fixes and is confidently wrong on 21%. Daytime imagery only. Not tested on the ESP32-P4 itself or on a real drone camera.

Licences and attribution. Code GPL-3.0-or-later. Trained on open geodata: © GeoBasis-DE/LGB (dl-de/by-2-0, modified); Geoportal Berlin TrueDOP 2022/2023 (dl-de/zero-2-0). The trunk was initialised from torchvision's ImageNet-pretrained EfficientNet-B0 (Apache-2.0 weights, BSD-3 code).

sha256 5be87d5e062d444fcd4ebc3bac3227d31d725b5d7bb05f8fb616cba4ca64bd9e