UAV-based crop and weed monitoring using RGB and multispectral imagery. Our pipeline spans orthomap generation and vegetation indices to deep-learning crop/weed detection and segmentation, including synthetic data generation for training.

Perception for self-driving vehicles, from LiDAR–camera calibration to 3D object detection and semantic segmentation. Our recent work adds real-time dual-task lane and drivable-area segmentation, plus continual domain adaptation for changing driving conditions.

Matching and retrieving 3D objects by shape, from full pottery scans to partial range-image queries. We combine panoramic-view descriptors, pose normalization and mesh segmentation for cultural-heritage and general-purpose 3D datasets.

Computer-aided diagnosis for breast cancer screening, using content-based image retrieval to give radiologists visually similar reference cases. Our SVM-based approach with a custom mHOG texture descriptor helps reduce unnecessary follow-up biopsies.

Detecting text embedded in natural scene photos — a much harder problem than scanned documents. Our approach uses cellular automata and bio-inspired models to localize text under real-world lighting, angles and clutter.

Distinguishing handwritten from machine-printed text in scanned documents, and spotting handwritten words across historical document collections — key steps for digitizing forms, archives and manuscripts.