AIT publishes world's largest image dataset for timber detection

At the prestigious IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) from 28 February to 4 March in Tucson, Ariz, researchers Julia Simon and Daniel Steininger from the Center for Vision, Automation & Control at the AIT Austrian Institute of Technology (AIT) presented for the first time the algorithm they developed together with their colleagues Andreas Trondl and Markus Murschitz for tree trunk detection in automated forestry, together with the largest publicly available image dataset to date.

TimberVision: Image dataset for the next generation of autonomous systems in forestry 

In forestry, many manual tasks such as inventorying, harvesting and measuring logs are not only time-consuming, but also require working in hard-to-reach or dangerous environments. The targeted use of automated machines and processes can help support workers and protect them from risks. This requires reliable technology that can reliably detect and measure logs and provide the recorded data for further processing. Until now, there has been a lack of sufficient training and reference data, which are essential for the development and validation of AI-based models.

This is where AIT's machine learning experts come in. To fill this data gap, the TimberVision project has developed a novel image data set for the robust detection of individual tree trunks and their exact contours. Based on this, the team has trained several different AI models and developed a fusion approach for the different results. Combined with other sensors, this will enable more efficient and automated inventory, as well as more precise harvesting and loading by automated machines.

The data was captured using standard monocular RGB cameras. A specially developed semi-automated processing pipeline is used for accurate and time-saving annotation. The AI was trained on the recorded data, which covered a wide range of conditions, including different locations, viewing distances, light and weather conditions, and tree trunk shapes and sizes. This variability was quantified using defined and measurable scene parameters. This allowed the scientists to ensure that the model was robust enough to reliably handle different environmental conditions and tree trunk appearances. The accuracy of the system was successfully tested in several trials. 

The world's largest publicly available image dataset for tree trunk detection

At the heart of this innovative project is an extensive, publicly accessible dataset of over 2,000 annotated RGB images and more than 51,000 recorded tree components, including cut and lateral surfaces. This is the largest collection of its kind.
To further advance research, the AIT team is making the entire TimberVision dataset and the algorithms developed publicly available for academic purposes. Scientists worldwide are invited to use and further develop the system.
 

The strengths of TimberVision at a glance 

• Largest available image dataset for tree trunk detection
• AI-based algorithms for accurate position and contour determination
• Separation and tracking of individual tree trunks across image sequences
• Geometric analysis to calculate centre lines and tree trunk dimensions for more precise handling
• Know-how transfer via GitHub for scientific collaboration 

With TimberVision, AIT is making a significant contribution to automated forestry through an easy-to-use system and a unique image dataset.
 

Reliable and accurate real-time detection, even in difficult weather conditions or when objects are obscured

Using state-of-the-art algorithms for object detection, segmentation and tracking, the system can reliably identify and recognise tree trunks and their components. Our work introduces a novel fusion and multi-object tracking framework to enable real-time detection of tree trunks, including their geometric properties such as outlines and centrelines. Our approach detects and segments the logs with a high degree of accuracy. The data is merged into a unified representation," explains Julia Simon, AIT software development expert. "What makes our system special is that it works reliably even under difficult conditions, such as bad weather or partial occlusion, and tracks the tree trunks precisely across image sequences," adds Daniel Steininger. He is an expert in AI and dataset development at AIT. With TimberVision, we are creating an important basis for the next generation of autonomous machines in forestry.
 

Links and further information

GitHub
github.com/timbervision/timbervision

Large-Scale Robotics Lab
www.ait.ac.at/labs/large-scale-robotics-lab

Paper
Daniel Steininger, Julia Simon, Andreas Trondl, Markus Murschitz. TimberVision: A Multi-Task Dataset and Framework for Log-Component Segmentation and Tracking in Autonomous Forestry Operations. Winter Conference on Applications of Computer Vision (WACV 2025).
https://arxiv.org/pdf/2501.07360

Press release (German)