WoodVision: Sim2Real Synthetic Image Generation for Classification of Wood-Jamming Events In Advance
2026 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
Student thesisAlternative title
WoodVision : Sim2Real baserad syntetisk bildgenerering för prediktion av produktionsstopp i ett sågverk (Swedish)
Abstract [en]
Sim2Real techniques offer a promising way to predict wood-jamming events in sawmill operations by bridging the gap between simulated and real environments. This thesis proposes and evaluates a Sim2Real pipeline that combines a physics-based 3D simulation, generative image enhancement, and machine learning. The pipeline is designed to enable early anomaly detection. The study investigates how simulation quality, diffusion-based realism improvements, domain-specific augmentations, and temporal modeling influence object detection and predictive performance in a real industrial sawmill. The key research questions focus on simulation realism, domain adaptation, and temporal anomaly prediction.The methodology begins by building a synthetic sawmill environment in Blender, where a rigid-body physics engine simulates realistic log-conveyor interactions. Logs are spawned with randomized poses and physical properties, creating a variety of plausible scenarios. Environmental and seasonal variations are introduced using NVIDIA Cosmos Transfer, a diffusion-based process that enhances image realism while preserving geometric and semantic consistency. Multiple datasets are generated by combining synthetic, diffusion-enhanced, and traditionally augmented images (e.g., HSV jitter, grayscale, blur). A YOLO-v11 model is trained for log detection and localization, with evaluation on real images to measure the reality gap. For temporal prediction, an MLP is trained on ground-truth object poses, and in inference, it uses predicted poses from the YOLO detector to anticipate jamming events.Results show that the synthetic-only model achieves moderate performance, with a pose mAP@50-95 of 57.76%, primarily limited by geometric accuracy and lighting conditions. Diffusion-enhanced images reduce the reality gap to a similar extent as traditional augmentation. Combining synthetic and real data significantly improves detection, reaching a pose mAP@50-95 of 88.41% and outperforming a model trained only on real images. The MLP, trained on annotated log poses, forecasts jamming events up to five seconds in advance with 94.9% accuracy and a 83.8% F1-score, following the hypothesis that abnormal angular deviations in log orientation precede wood-jamming events. This demonstrates that temporal modeling can effectively anticipate short-term anomalies. These results emphasize the importance of accurate simulations, consistent lighting, and careful dataset design for effective Sim2Real transfer. Specifically, experimental comparisons reveal that data quality and geometric fidelity are more critical than sheer quantity, as scaling synthetic data can increase model bias toward the simulated environment. The thesis provides a practical framework that integrates simulation, generative image enhancement, and temporal modeling to support early anomaly detection. Future work should explore higher-detail simulation models to better capture log dynamics and lighting, test advanced domain adaptation and augmentation strategies, and validate temporal predictions on larger real-world datasets. These improvements are expected to further reduce the reality gap, enhance predictive performance, and strengthen the practical use of Sim2Real pipelines in industrial sawmill operations.
Place, publisher, year, edition, pages
2026. , p. 104
Keywords [en]
Sim2Real, Reality Gap, Diffusion, 3D Simulation, Object Detection, YOLO, Temporal Anomaly Prediction, Sawmill, Wood Jamming
National Category
Computer Vision and Learning Systems
Identifiers
URN: urn:nbn:se:liu:diva-224536ISRN: LIU-IDA/LITH-EX-A--26/029--SEOAI: oai:DiVA.org:liu-224536DiVA, id: diva2:2067641
External cooperation
AFRY
Subject / course
Computer Engineering
Presentation
2026-05-28, 3F:469 (Alan Turing, E-huset, Campus Valla, Linköping Universitet), Linköpings universitet 581 83 Linköping, Linköping, 16:00 (English)
Supervisors
Examiners
2026-06-222026-06-082026-06-22Bibliographically approved