Inside a 12.4 km audit: reading 128 defects on NH-44

A defect list is not a finding. Handed 128 rows of detections, most people read the total, wince, and file it. The value is in the four or five questions you ask of the rows afterwards — and those questions are the same whether the corridor is this one or yours.

The run

One vehicle, one pass, a standard dashcam recording 1080p at 30 fps. Chainage 428.0 to 440.4 on the Hyderabad–Kurnool carriageway: 12.4 kilometres, driven at normal traffic speed with no lane closure and no second visit. The output is 128 detections across eight classes, each carrying a severity grade, a confidence score, a coordinate and a source frame.

Question one: what is actually broken?

Sorted by class, the corridor is overwhelmingly a pavement problem rather than a roadside-furniture problem. Potholes, cracking and ravelling account for 83 of the 128 detections; barriers, kerbs and signage together account for 24. That single split determines which department owns the report and which budget pays for it.

Three quarters of everything found is surface distress. That single fact already routes the report: this is a works-department problem, not a safety-furniture procurement one, and the two live in different budgets.

The 21 faded lane markings deserve a second look, though. Marking visibility is cheap to restore and disproportionately tied to night-time and wet-weather risk — it is usually the highest safety return per rupee on a list like this, and it is almost always the item that gets deferred because nothing about it looks urgent in a photograph.

Question two: how much of it matters this month?

Severity splits 51 low, 51 medium and 26 high. Only the 26 need action now — a fortnight of crew time, not a resurfacing programme. Within them the concentration is sharper still: 21 of the 26 urgent items are just two failure modes on the running surface, potholes and alligator cracking.

Twenty-six high-severity defects across 12.4 km is a fortnight of crew time, not a resurfacing programme. And within that 26, the concentration is sharper still — 14 are potholes and 7 are alligator cracking, so 21 of the 26 urgent items are two failure modes on the running surface. The remainder is a scatter of kerb, crack and signage issues.

Meanwhile the 51 low-severity findings are not this year’s problem at all. They are the baseline you compare against on the next drive, which is where the value of repeat surveys shows up: a low-severity crack that is still low-severity in three months can be ignored with confidence, and one that has progressed cannot.

Question three: where is it?

Binned by kilometre, the corridor is unremarkable except in one place: 57 of 128 detections — 45% of the entire defect load — fall inside the two kilometres between chainage 435 and 437, which is 16% of the length. Everything either side is comparatively healthy. That pattern is the finding.

Binned by kilometre, the corridor is mostly unremarkable — between two and eleven detections per kilometre across most of its length. Except here:

Fifty-seven of 128 detections — nearly half the corridor’s entire defect load — sit inside a two-kilometre window that is 16% of its length. Everything either side of it is comparatively healthy.

That pattern is the finding. A uniform scatter of 128 defects over 12.4 km would describe a corridor at the end of its service life, and the answer would be a resurfacing bid. A cluster this tight describes a localised cause — drainage, a subgrade weakness, a construction joint, a junction where heavy vehicles brake and turn — and the answer is a site investigation of two kilometres, at a fraction of the cost.

A ranked list sorted by severity would have scattered those 57 defects through the document and hidden this completely. It is only visible because every detection carries a coordinate.

Question four: can I trust any individual row?

Mean confidence across the 128 detections is 0.91 — healthy, but not a licence to skip review. Treat the confidence column as a triage dial rather than a grade: filter high when building a capital case, drop the threshold when investigating a specific stretch, where a missed pattern costs far more than a false positive.

Concretely: filter high to build the capital case, then drop the threshold and look at everything when investigating that 435–437 window, because in an investigation a false positive costs you a glance and a false negative costs you the pattern.

Each row also carries the frame it came from, so any detection can be checked visually in seconds. That is what makes a disagreement with the model resolvable rather than a matter of opinion — and it is the property to insist on from any tool in this category, as we argue in the piece on reading accuracy figures.

What the report actually recommends

Read together, the four answers produce a plan with an order to it: investigate the 435–437 cluster before repairing anything in it, programme the 21 urgent surface defects elsewhere as routine patching, restore the 21 faded markings as one cheap line item, then re-drive in 30 days and compare.

1. Investigate km 435–437. Not repair — investigate. A cluster this dense has a cause, and patching the symptoms of an unresolved drainage or subgrade problem buys one season.
2. Programme the 21 urgent surface defects outside the cluster as routine patching.
3. Restore the 21 faded markings as a single low-cost line item, ahead of the monsoon rather than after it.
4. Re-drive in 30 days and diff the result. The 51 low-severity findings become a change signal instead of a backlog.

None of that required an engineer to walk 12.4 kilometres. It required one drive, and then an engineer’s judgement applied to two of them — which is the whole argument for this way of working, set out at greater length in the manual-versus-automated comparison.

Look at the data yourself

Every number above is computed from the dataset in the public console, which is unlocked and exportable. Filter it by class or severity, open the map to see the 435–437 cluster in position, and download the CSV to check the arithmetic. It is the same console, and the same export, a live audit is delivered in — see the product overview for the full list of deliverables.