Model card Road damage detector

How the model sees
and where it doesn't yet

A plain account of what Poth's detector was trained on, how it scores, and its limits. Version 0.2 learned from 8,834 more Indian road frames and now scores higher on Indian roads it never saw.

Version
v0.2 · fine-tuned on Indian roads
Architecture
RF-DETR Large, detection transformer, 33.6 M parameters
Input
A dashcam or phone frame, resized to 704 px
Output
Boxes in 4 damage classes, each with a confidence score
Serving
Live API on Modal · about 1 s per frame on CPU
Updated
14 November 2026
46.6%mAP@50+1.4 vs v0.1box overlap ≥ 50%
21.0%mAP@50:95+0.7 vs v0.1strict box overlap
58.3%Precision+1.2 vs v0.1share of flags that are real
50.1%Recall+0.1 vs v0.1per frame; repeats across frames raise it
53.9%F1+0.6 vs v0.1balance of the two
61.7%Seven-country mAP@50-0.8 vs v0.1the original validation set

Locked Indian test set, 1,150 frames never used in training · v0.2 against v0.1 · threshold 0.35

What v0.2 changed.
More Indian roads, scored honestly.

We fine-tuned v0.1 on 8,834 new Indian frames from two public sets, with the Indian training set shown twice as often. Frames close to any test frame were removed first, so the test stays unseen.

+6.6 points on transverse cracks, the hardest class

+1.4 points overall on the Indian test

−0.8 on the seven-country set, a small trade for local accuracy

Indian test, by damage type

v0.1v0.2
  • Longitudinal crack41.4%no change
  • Transverse crack28.3%+6.6
  • Alligator crack67.2%-2.2
  • Pothole49.4%+1.2

mAP@50, 1,150 held-out Indian frames

One frame, scored live.
And how training went.

A frame the model never trained on

People's labels, then the live model's output

Dashcam frame with six defects labelled by peopleD40D40D40D40D20D00
Labelled by people 4 potholes, 1 alligator crack, 1 long crack
The same frame with the model's detections80%74%71%59%46%39%
Found by the model all 4 potholes; the 2 cracks located but called potholes

Training curve

Validation score after each epoch

45505560650481216
mAP@50F1Epoch · best saved at 16

Rose fast, then flattened after epoch 12

Four kinds of damage

Labelled defects in training

  • Longitudinal crackD0026,016
  • Transverse crackD1011,830
  • Alligator crackD2010,616
  • PotholeD406,544

55,006 labels on 38,385 images, plus 24,946 Indian labels in v0.2

A test set kept apart

Indian frames, split in blocks of 50

1,150 Indian frames

Used once, for the final score

Public data today.
Assam's own next.

Today's scores come from a public, multi-country dataset, so there is real room to improve. With investment, we will build a road-damage dataset native to Assam, giving higher scores on the roads that matter and a result agencies can trust.

v0.1Retired

Public data, seven countries. Indian test mAP@50 45.2%.

v0.2Live now

Fine-tuned with 8,834 more Indian frames. Indian test 46.6%.

v1.0With investment

Fine-tuned on an Assam dataset, scored on Assam roads it never saw.

1,000+ km driven across Guwahati, Dibrugarh, Tinsukia and rural roads

Every season dry, monsoon-wet and patched surfaces

Labelled with engineers using the same four damage classes

A locked Assam test set split by road stretch, scored once, published

Training data: RDD2022, Arya et al., 2024, CC BY 4.0 · Base model: RF-DETR (Apache-2.0)

Built for
and not for.

Intended use

Screening road networks for cracks and potholes from routine drives

Prioritising inspection so engineers start where damage is worst

Checking repairs by comparing one drive with the next

Out of scope

Final decisions without review. Low-confidence finds go to an engineer.

Identifying people or vehicles. Faces and plates are blurred before storage.

Night, flooded or snow-covered roads. Too few training examples to trust.

Known limitations

Few Indian cracks yet. v0.2 added 8,834 Indian frames, but 95% of their labels are potholes, and none come from the North East.

Cracks get called potholes. On close-up damage the classes blur; see the frame above.

Thin transverse cracks are hardest. They are the smallest class in the data and easily confused with joints and shadows.

Scores are per frame. A defect seen across several frames is far less likely to be missed than recall suggests.

Put it on your next drive.
We'll send back the list.

Share one route's footage. Get every defect on it, boxed, classed and mapped.