Selected Cases

LogStack is a suite of digital measurement tools for raw timber stacks in the forest, comprising LogStackPRO, a certified handheld device, and LogStackLIDAR, an app for compatible iPhone Pro and iPad Pro devices. Both reconstruct stack volume, diameter distributions and log positions without manual reference measurements, storing results in the cloud. Talmo has developed the full product line, from prototypes through production software, for HedeDanmark, before it was later acquired by the Swedish company ForestX.

LogStackPRO is a Linux-based system built around a custom stereo camera, with Talmo devising the camera design and calibration to reconstruct the metric 3D structure of the stack. LogStackLIDAR instead builds on Apple's ARKit, running on iPhone Pro and iPad Pro hardware. Both products share the same underlying technology stack, relying heavily on simultaneous localization and mapping (SLAM), photogrammetry and deep learning-based detection and segmentation for precise stack density and volume estimation. The reconstructed depth data also enables per-log measurements and a visual 3D representation of the stack, which can be used to verify the automatically generated results.

Talmo has taken the products from research prototype to certified, field-ready software. Today, LogStack measures more than 2.5 million cubic meters of timber annually across Denmark, Germany, Hungary and Poland, giving foresters, hauliers and mills better control over volumes, quality and logistics.

For RealHorse, Talmo has developed prototype apps and the underlying analysis engine for app-based horse lameness detection, diagnosing lameness from videos recorded on hand-held devices without the need for specialized equipment. The goal is to let riders, trainers and veterinarians catch subtle movement asymmetries early, before they develop into more serious lameness issues. The system's accuracy has been validated in a peer-reviewed study published in the Equine Veterinary Journal, comparing results from thousands of strides against a research-grade 3D reference system.

To support this, Talmo has developed a deep learning model for real-time key-point detection of the horse's limbs and body, tailored specifically to the gait analysis use case. Since no suitable annotated dataset existed, Talmo has also built a custom semi-automatic annotation tool to efficiently label the large volumes of video footage needed to train and validate the model.

The detected key-points feed into a frequency-based signal analysis pipeline that extracts stride features such as symmetry and timing, which are then used to generate an automated diagnosis. The prototype was implemented as an iOS Swift app, combining on-device CoreML inference for the deep learning model with a custom C++ processing engine for the signal analysis, enabling fast results directly on the user's phone.

Talmo has helped Terma develop video-based drone detection software for observing and tracking intruding UAVs in high-end surveillance cameras for deployment at critical infrastructure sites.

The solution performs registration of the video stream for continuously learning the camera pose, along with employing different deep learning-based detectors trained on both real and synthetic data. Finally, the system performs object tracking based on motion models and Kalman filtering.

The tasks performed involved evaluating different detector architectures with respect to both their training and inference capabilities for deployment in a large system with many parallel camera streams. Moreover, the tasks involved devising ideas for fusing radar and camera tracks.

That work supported Terma's counter-drone product line for sites that need reliable detection across many simultaneous camera feeds, not a one-off lab demo.

For Kozoom Carom Billiard, Talmo has developed a desktop application that is capable of analyzing and generating 4k video at 60 fps captured through Blackmagic DeckLink equipment.

The application detects the Carom table, tracks the three balls and freezes the exact moment when the shot is made. The analyzed results are overlaid on the input video and output through the Decklink card for mixing with other camera sources for visualizing details after the shot has been made. It also enables the video operator to control which balls are tracked and adjust the graphics overlay.

The video analysis operates at 60 fps hence requiring highly optimized routines for video input processing and scaling, color space and bit depth conversions along with the actual object tracking.

Christmas tree growers traditionally assess their plantation stock by manually counting trees in a small sample of the plantation, typically around 20%, and dividing them into height and quality classes before extrapolating the result to the full stock. Talmo has been part of a consortium along with University of Southern Denmark (SDU), Hermes Traffic Intelligence and Green Team Group exploring how drones can be used to give a status of the Christmas trees in a given plantation including estimating the number of Christmas trees and the heights of the individual trees.

Images acquired from an off-the-shelf drone are used to generate an orthomosaic, height model and matching 3D point cloud of the plantation, which are then interpreted to localize the individual trees and produce statistics for the full plantation. Talmo has supplied knowledge about photogrammetry, object detection and segmentation along with 3D data analysis to the project.

Together with brake pad manufacturer SBS, Talmo has worked on prototypes for detecting and identifying spare parts from images. Motor-cycle brake pads come in many shapes and sizes, where each brake pad is compatible with only a certain subset of end-user braking systems.

To leverage existing SBS CAD-based data for each known brake pad, Talmo has built a rendering solution, where multiple synthetic images can be generated with random wear and tear for each model, from various viewpoints, lighting setups and background scenes.

The generated images can be used for training both detection and segmentation systems without the need for manually acquiring and annotating millions of images to support all combinations. The trained inference models can be used both for factory quality control, but also for end-user apps capable of identifying the actual brake pads from phone images, while finally providing suggested replacement model numbers.

The prototypes showed that CAD-driven synthetic data can cover a large pad catalogue without photographing every SKU by hand, for both factory QC and end-user identification.