This work was carried out at Timbeter, where Martin Kambla was a co-founder and CTO, before Kratt Ventures Limited existed. It is presented here as an example of the founder’s hands-on experience. It is not KRATT client work, and Timbeter is not a KRATT client.
The challenge
Timber is traded by volume, and stacked logs are traditionally measured by hand. The technical goal was to detect and measure individual log faces from ordinary photographs, taken on a phone, in real forest, roadside and port environments.
Real conditions made this hard. Photographs vary in lighting, weather, angle, distance and bark condition. Log ends can be dirty, shadowed, snow-covered or partly hidden by neighbouring logs. A method that worked on clean test images was not enough. It had to hold up in the field.
Early investigation
The first approaches used traditional image processing and geometric methods: edge and shape analysis, and fitting geometric models to candidate log faces. This established what could be done with explicit rules and, just as importantly, where rule-based methods broke down in real-world imagery.
Machine learning
Moving to learned detection meant building labelled datasets from real field photographs, training detection models and evaluating them systematically against held-out data. The dataset grew over years of real-world use. That growth, and the iterations on the model it allowed, were central to the progress.
From detection to measurement
Detecting a log is not the same as measuring it. The measurement pipeline added segmentation and boundary analysis of each log face and converted image-space results into physical measurements. The team then evaluated how errors at each stage carried through to the final figure.
Neural-network evolution
As the field matured, the detection approach moved towards modern neural-network methods. Each change of approach was judged by measured results on representative data, not by novelty.
Inventorship
Martin is a named co-inventor on a published international patent application for image-processing methods in this domain: WO2017114977A1.
Later validation work
Later work included stereo-depth measurement and the technical validation needed for Lithuanian type approval of the measurement approach. This is the kind of evidence a measurement system needs before people rely on it commercially.
Capabilities demonstrated
- Computer vision and image processing in uncontrolled real-world conditions
- Machine learning, from dataset development through systematic evaluation
- Experimental engineering: forming and testing candidate approaches over time
- Measurement-system design and technical validation
- Taking research results into a production product used in the field