Kerb RDD Nano β on-device road-damage detector
RF-DETR Nano fine-tuned on RDD2022 (Japan-labelled training set),
exported to ExecuTorch .pte with XNNPACK delegation for
react-native-executorch on-device inference. Part of the
Kerb privacy-preserving
road-inspection system β detection runs on the phone, only damage
crops leave it.
Artifacts
| File | Size | Schema |
|---|---|---|
kerb_rdd_nano_xnnpack.pte |
112 MB | canonical β wrapped RN-ET contract: input [1,3,512,512] f32 ImageNet-norm -> boxes[N,4] xyxy-normalized, scores[N], classes[N] f32 (0-indexed argmax; label = label_map value index+1) |
kerb_rdd_nano_raw.pte |
108 MB | raw RF-DETR export: dets[1,300,4] cxcywh + labels[1,300,10] logits β fallback for custom runtimes |
sha256: wrapped ddf024c4β¦faaf4d Β· raw 0c57651eβ¦a7e055f
(full hashes in export_meta.json)
Verified results (Kaggle v5 train + v3 export, 2026-11-17)
mAP50:95 = 0.3711 (best EMA, Japan train)
Domain shift β golden mAP@50 per region:
JP US China MB Czech China Drone India Norway 0.6907 0.6388 0.3185 0.2618 0.2763 0.2040 0.1443 Runtime-verified:
scripts/pte_smoke.pyloads both.ptefiles in the portable ExecuTorch runtime + XNNPACK and executes a forward pass (3-output wrapped schema confirmed end-to-end).
Classes (9; model output index i β label_map value i+1)
Block crack (0) Β· D00 longitudinal crack (1) Β· D10 transverse
crack (2) Β· D20 alligator crack (3) Β· D40 pothole (4) Β·
D43 crosswalk blur (5) Β· D44 white line blur (6) Β·
D50 utility hole (7) Β· Repair (8) β see label_map.json.
Use
// react-native-executorch β₯0.11 (legacy URI API)
import { ObjectDetectionModule } from 'react-native-executorch/legacy';
const m = await ObjectDetectionModule.fromCustomModel(source, false);
const dets = await m.forward(frameUri); // boxes/scores/classes
Or the portable runtime:
from executorch.runtime import Runtime
m = Runtime.get().load_program("kerb_rdd_nano_xnnpack.pte").load_method("forward")
boxes, scores, classes = m.execute([tensor_1x3x512x512])
Limitations
- Trained on Japanese roads β large domain shift elsewhere (table above). Per-region adaptation is milestone M4 of the project.
- Scores are not calibrated probabilities; treat
confas a ranking signal until Mondrian calibration lands (M4). - Detects road damage only β it is not a face/plate privacy model.
Provenance & license
Fine-tuned from rfdetr-nano (Roboflow, Apache-2.0) on RDD2022
(Arya et al., IEEE Big Data 2022) β CC BY-SA 4.0. These weights
are a derivative work shared under the same license. Cite both:
Arya et al., "RDD2022: A Multi-Country Road Damage Dataset",
IEEE Big Data 2022. https://github.com/sekilab/RoadDamageDetector
- Downloads last month
- 42