Showcase
All test photos
Your photo
A small model runs in this browser: the photo never leaves your computer (first use downloads 27 MB).
Model
Houses
Entry point
Rescue team enters:
Legend
- Reachable by truck (confident)
- Probably by truck: check first
- Boat needed
- Last stretch on foot
- No route: air team
- Truck Boat Foot
- Route on the other map, when they disagree
- Entry point (truck or boat)
About
A drone photo is mapped into five classes by one of two models trained on the same 50 labelled photos (FloodNet) or 20 (RescueNet), with 30 more labelled validation photos per dataset for the decode thresholds and the router settings:
Chameleon, the few-shot model this project started from, and a standard segmenter (an ensemble of three
DeepLabV3-ResNet50 with flipped views). Each dataset opens on the model that is better there (see
below); the Model switch shows the other. A router on 4×4-pixel cells then finds the cheapest way from the entry point to each house: truck on dry ground
(roads, or yards at 4× cost; never where water is likely), then a boat across water, then the last stretch on foot
(never back into a vehicle). Vehicles have a width: a truck needs about 2.5 m between buildings and a boat
about 1.6 m, so neither squeezes through the gap between two houses; people on foot can. The labelled map is the same router run on the human labels: the datasets label pixels,
not reachability.
Confidence: the prediction is routed twice, with the normal off-road rule (land where p_wet < 0.3 for
FloodNet, 0.2 for RescueNet, each tuned on the dataset's validation photos) and a cautious one (p_wet < 0.05 /
0.15). Houses a truck reaches even on the cautious map are confident (on FloodNet test photos 99 % truly
truck-reachable for the ensemble, 100 % for Chameleon); the others are marked "check first" (82 % / 79 %). On the
full FloodNet test about 7 % of the houses the ensemble shows as truck-reachable were not (Chameleon 11 %), mostly
where the route crosses water the model missed. Minutes and metres are
rough estimates (truck 20 km/h on roads, 5 off-road, boat 6, on foot 3; scale from the cars in the photo).
FloodNet, full 448-photo test: the DeepLabV3 ensemble mIoU .62, Chameleon .51 (PSPNet trained on all 1,445
labelled photos: .85 on these classes). A pre-registered comparison on the same photos (Phase 2) found Chameleon
ahead only with very few labels (1-5 FloodNet photos) and on a new disaster at 10-20 photos, where its 30 tuning
photos help: RescueNet here, 20 photos, mIoU .42 against .37 for the DeepLabV3 ensemble (eval100), so RescueNet
opens on Chameleon. Both models use the same router and gates, each with its own statuses, tiers and per-photo IoU.
Houses the other model routes differently are marked in the house list and counted under the map; the sort
"models disagree most" lists those photos first. One showcase scene shows an honest failure of Chameleon: shallow
water read as dry land.
Besides the 8 showcase scenes you can browse all 200 test photos of the evaluation split (100 FloodNet,
100 RescueNet), not picked by us: an orange count is the houses a truck cannot reach on the prediction, a pink dot
marks a photo where prediction and labelled map disagree. "Water probability" shows what the model believes
cell by cell. The address bar keeps the scene, house, entry point and view, so a link opens exactly this view.
"Try your own drone photo" maps a photo in this browser with a small model
(LR-ASPP-MobileNetV3, 13 MB, trained on the same 50 labelled FloodNet photos plus unlabelled ones the baseline
labelled itself: full-test mIoU .60) and routes
it with the same rules; the photo is never sent anywhere. It has no labels, so the houses and the entry point come
from the model's own map; it suits drone photos taken straight down at FloodNet's height.
Source code, the evaluation splits and how every number was measured:
github.com/MarlenMM/rescue-router.
A research demonstration, not a tool for operational rescue decisions.