Deep Learning based detection and correction of print defects like missing nozzles in real-time, without any setup or operator intervention. A custom CNN detects defect position and severity. This enables a closed loop where print quality can be continuously assessed and adjusted in real-time.

We developed a stereo vision system powered by AI accelerators and custom CNNs for accurate 3D localization of a printed color chart. The setup enables precise robot control and reliable hand-eye calibration. This know-how supports high-precision automation tasks.

Our expertise in data balancing for semantic segmentation enhances deep learning models by dynamically rebalancing class distributions during training. Through our published work on Dynamic Label Injection (DLI), presented at the VISION Workshop at ECCV 2024, defects are seamlessly transferred using Poisson-based image cloning and cut-paste techniques, ensuring more effective learning. This approach improves model performance, even in challenging weakly-supervised settings.

Our expertise in synthetic data generation enhances industrial vision models by leveraging diffusion models conditioned on simple annotations such as bounding boxes. Through our published work, presented at the SynData4CV Workshop at CVPR 2025, we generate high-quality, diverse datasets with minimal manual effort, enabling more accurate and robust semantic segmentation. This approach maximizes label efficiency while scaling data generation to meet complex industrial needs.

We deploy efficient, production-ready deep learning solutions on edge devices, leveraging AI accelerators for low-power, high-performance inference. Our expertise spans neural network optimization, quantization, and pruning — techniques that compress and accelerate models to boost efficiency without sacrificing accuracy — across SoCs like Hailo, Qualcomm, SiMa, and Axelera.

We developed expertise in designing and optimizing multi-camera setups for complete surface coverage and minimal occlusion. The project includes calibration, synchronization, and layout optimization to ensure high-quality, consistent imagery. This know-how supports accurate inspection and 3D reconstruction in industrial settings.

We specialized in high-precision 3D reconstruction, combining advanced imaging technology with state-of-the-art methods. Using high-resolution industrial cameras, tailored illumination setup, structured light, and high-bandwidth networking, we capture data that is processed by our methods, which leverage differentiable physically based and neural rendering to create highly detailed and realistic 3D models. Our expertise in 3D reconstruction contributed to the creation of the Covision Media spin-off.