Skip to main content
    Computer Vision & Drone Inspection

    Drone Inspection Software & Computer Vision Development - Defect Detection From Aerial Footage

    Drones made capturing inspection footage cheap. Reviewing it didn't get cheaper - an inspector still watches hours of video to find a handful of defects. We build computer-vision pipelines that surface the candidates, let your inspectors make the call, and put every confirmed finding on the right asset.

    See case studies
    From flight to confirmed finding
    1. 1
      Ingest
      Video, stills and GPS / altitude telemetry per flight
    2. 2
      Sample & dedupe
      Distinct views only - overlap removed
    3. 3
      Detect
      Detection & segmentation per defect class
    4. 4
      Inspector review
      Confirm, reject or correct each detection
    5. 5
      Geotag & report
      Findings on the map, the asset and the report

    Reviewer decisions feed back into training - the model improves as part of normal inspection work.

    What you get

    Ingestion for drone video, high-resolution stills and flight telemetry (GPS, altitude, heading) - batched per inspection, not per file
    Frame sampling and near-duplicate removal tuned to flight speed and altitude, so models and people only see distinct views
    Defect detection and segmentation models (YOLO-family detectors, segmentation networks) trained in PyTorch on your imagery and exported to ONNX
    Labeling workflow with active learning - the frames that would teach the model most are labeled first, so inspector time goes further
    Human-in-the-loop reviewer interface where inspectors confirm, reject or correct every detection before it is reported
    Geotagging of confirmed findings onto a map and your asset register, plus draft report generation from structured findings
    Evaluation set with precision and recall tracked per defect class, gating every model release
    Deployment plan covering cloud GPU batch inference after each flight and edge options (on-drone or ground station) where results are needed on site

    When it fits

    • You already capture drone imagery at volume and inspector review time is the bottleneck
    • You can name the defect classes that matter (corrosion, cracks, missing components, vegetation encroachment, hotspots) and what each one looks like
    • Your inspectors can give a few hours per week to labeling and review in the first months - the model learns from them
    • Findings need to land in an existing asset register, GIS or reporting workflow rather than a standalone dashboard

    When it doesn't

    • You need a drone operator - we build the software; your pilots or flight service provider fly the missions under Part 107, EASA or local rules
    • Fully automated reporting with no human sign-off is a requirement - we design for inspector confirmation on every reported defect
    • The imagery doesn't exist yet and the capture protocol (altitude, angle, overlap, sensor) isn't settled - fix capture first, or models will be trained on inconsistent data
    • An off-the-shelf inspection platform already covers your asset class and defect types - license it and integrate it instead

    Process

    Weeks 1–3: discovery and data audit - sample flights, defect taxonomy, capture-protocol review, and a labeled evaluation set agreed with your inspectors. Weeks 4–8: ingestion, frame sampling, a baseline model for the first one to three defect classes, and the reviewer interface, so inspectors start confirming real detections early. Weeks 9–14: active-learning rounds, geotagging and asset-register integration, report generation, and production batch inference. Edge deployment and additional defect classes follow as separate phases, each gated on the evaluation set.

    Full delivery process

    FAQ

    How much labeled data do we need to start?
    Less than most teams expect, but more than zero. For a visually distinct defect class, a few hundred well-labeled examples is usually enough for a useful first model; subtle or rare defects need more. We start from pretrained detection backbones, label the most informative frames first with active learning, and use inspector review decisions as new labels - so the dataset grows during normal operations instead of in a separate labeling project. The data audit in discovery tells you how much you already have.
    Can the models run on the drone or at the edge?
    Often, with trade-offs. Detection for a small number of defect classes can run on Jetson-class hardware on the drone or in a ground station, which is useful when a crew needs results before leaving the site. Full-resolution analysis across many classes is usually better as cloud GPU batch inference after the flight. Models are exported to ONNX so the same model can be deployed to either target; we measure accuracy and latency on your actual hardware before recommending one.
    How accurate will it be?
    We won't quote a number before seeing your imagery. Accuracy depends on the defect class, image resolution, capture consistency and label quality. What we do commit to is measurement: a held-out evaluation set agreed with your inspectors, precision and recall reported per defect class, and confidence thresholds chosen with you - for safety-relevant defects you usually trade more false alarms for fewer misses. Every reported finding is still confirmed by an inspector.
    Who owns the trained models and the data?
    You do. Imagery, labels, trained weights, training code and evaluation sets are assigned to you under a standard work-for-hire agreement. We also check the licences of any pretrained models and libraries up front - some popular detection frameworks carry copyleft or commercial licence terms that matter if the model ships inside your product.
    How does this fit our existing inspection workflow?
    It slots in between capture and reporting. Your pilots fly the same missions; footage is uploaded as it is today; inspectors work in the reviewer interface instead of scrubbing video; confirmed findings are pushed to your asset register, GIS or maintenance system through its API or an export your team already uses. We map the current workflow in discovery and change as little of it as possible.
    TekPiq engineers in an architecture session at the whiteboard, one of them sketching the data model
    Architecture session in our office - client details on the board are blurred.

    Who builds it

    Our own team, working from our office in East Sarajevo, Bosnia and Herzegovina. The engineers on your scoping call are the ones who build it.

    Meet the team

    Ready to talk computer vision & drone inspection?

    30-minute scoping call. No obligation, no hard sell.