What has computer vision achieved in log yards?

Accurate measurement where the images are controlled, and weak detection where they aren't. A 2024 study that photographed log ends with reference markers measured a whole truckload's volume within 1.20% and single diameters within 0.73% in a forestry test. Another 2024 study built a machine-vision sorting line for small logs, with diameter errors averaging 1.12 mm and sorting accuracy above 95%. By contrast, a low-cost 2025 model working from ordinary forest and trailer photographs found under 60% of the logs in its test images and reported no measured diameter error. We read the two 2024 studies from their abstracts.

What the usual answer says

Searching the exact question returns pages about computer vision in container and truck yards: licence plates, trailer IDs and safety. Log yards barely appear. The general claim is that vision counts, measures and scales logs. The research supports that for some set-ups and not others.

Controlled capture: log ends and sorting lines

Lu and colleagues, in Forests in 2024, combined two networks to find and segment log ends in photographs. They added a visual marker system "to estimate the camera position during image acquisition", reducing distortion from the shooting angle. In a forestry test, "the measurement errors for the volume of an entire truckload of logs and a single log diameter are 1.20% and 0.73%, respectively". The authors say these are within the industry standards they used.

Ding and colleagues, also in Forests in 2024, built a sorting line where logs pass fixed cameras. The paper reports that "The absolute error in diameter detection for the sorting line averages 1.12 mm". Sorting accuracy exceeds 95%, and the line handles logs from 60 to 300 mm across, at 120,000 to 130,000 cubic metres a year.

Uncontrolled capture: ordinary photographs

A July 2025 preprint by Hasanzadeh Fard, Hasanzadeh Fard and Jonoobi set out to estimate "timber log diameter using standard RGB images captured under real-world working conditions", without "expensive sensors or controlled environments". The model was fine-tuned on a subset of a public dataset whose images come from dashcams in forests, at roadsides and on trailers. On a test of 21 images with 208 logs, it reached "a precision of 0.656, recall of 0.577", so it found under 60% of the logs present. On the stricter localisation measure, mAP@0.5:0.95, it scored 0.356.

The paper contradicts itself on this. Its results call the performance "strong" and describe "minimal false positives or missed detections", beside a recall that misses four logs in ten. Its methods promise diameter "bin accuracy" against manual labels, but the results report none. Diameter ranges only "appeared visually consistent".

Better-resourced work on ordinary imagery is growing. TimberVision, released in 2025, has "more than 2k annotated RGB images containing a total of 51k trunk components", aimed at forestry machines rather than yards.

Reading it through a data lens

The pattern is about capture. Where the camera sees log ends at a known pose, or logs pass a fixed station, measurement reaches errors near 1%. Where it sees piles and trailers from wherever the camera happens to be, detection itself is unreliable. The reason is scale. A marker or a fixed station tells the model how many pixels make a centimetre. An ordinary photograph doesn't, and the 2025 authors note that box-based diameters "can be affected by camera angles, overlapping objects, or inconsistent scaling, which may reduce accuracy in real-world use". Better models don't remove that problem; better capture does. A 2024 review of 33 studies adds a warning: "comparing performance across studies proved challenging due to varying goals and metrics". A yard should compare any system on its own logs, with measured error.

Where it touches an operation

The check is reconciliation. Where scale tickets carry piece counts and volumes, a camera system's counts and volumes can be set against them load by load. Differences show whether the system misses logs or mismeasures them, and on which loads.

What we cannot take from it

We read the two 2024 studies from their abstracts. Both used conditions of their own choosing, small logs in one case, and neither result transfers directly to a different yard, species or camera. The 2025 preprint is a single small test.

Quarri for sawmills is built around how a sawmill runs, from log intake to shipped order.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.