Services

From prototype to production

Our expertise from computer vision, machine learning and applied maths extends far beyond traditional image processing, enabling us to tackle a wide range of challenges. Whether you're working with images, video or entirely different data, we deliver innovative, tailored solutions to meet your unique needs.

Recognition problems: Semantic Segmentation

At the core of computer vision lies the task of recognizing objects and structures in images. Depending on the desired output granularity, these problems are called image classification, object detection or semantic segmentation. The data can RGB images, Lidar data, video or three-dimensional medical images. In any case, the solution is normally based a convolutional neural network (CNN) or a vision transformer that is learnt from data. If you call it a machine learning, deep learning, artificiial intelligence or even a filterbank is up to you,.

3D reconstruction: SLAM and SfM

3D reconstruction is the task of estimating a 3D model of a scene from 2D images or video. The first task is typically to estimate the pose of the camera for each image or frame. Given these poses, points seen in multiple images can be triangulated to estimate its 3D position. The result is a 3D point cloud model with camera poses and 3D points. Depending on your background and the exact setup, this process is called photogrammetry, structure from motion (SfM) or simultaneous localization and mapping (SLAM).

3D reconstruction: Gaussian splatting

Structure from motion and SLAM produces sparse point cloud models of the scene. Sometimes are more complete model is needed, for example to analyze the shape of objects in 3D. A modern approach to this is called Gaussian splatting. The points in the sparse point cloud model are replaced with Gaussians. The position, color and opacity of these Gaussisans is optimized to fit the 2D images. Gaussians are removed or split when required often resulting in a very accurate representation of the scene.

Image matching and alignment

A special type of recognition problem is image matching, being the process of identifying corresponding points or structures between pairs of images. This is often the first step in identifying the geometric relationship between the images, for example to align different image modalities, or for stitching images to a panorama.

Multi-view object detection in 3D

Given the poses of a set of cameras in 3D, this is the task of accurately detecting objects in 3D. Previously, detection was typically performed in 2D and detections from different views were tracked or merged using 3D information, but more recently AI-based

Human pose estimation

A special case of object detection, human pose estimation consists of detecting the position of key structures of the human body. Often a step in analyzing the actions of a person.

How can we help you?