ML-based Computer Vision Models

AMDC has been actively working on developing ML models to perform image classification and object detection for various applications.

Image classification

In Computer Vision the term Image Classification refers to the process of assigning a certain class to an image based on the visual content.

AMDC customly develops task-specific Image Classifiers based on the latest international research results. Our particular focus is on achieving extremely low execution time on mobile embedded hardware, such as NVidia Jetson or the latest FPGAs.

Classification of a drone within a Region of Interst

Object Detection

Object Detection is the process of localizing and classifying certain objects within an image.

AMDC Object Detection is based on state-of-the-art Object Detection Model Architectures, such sas YOLO, SSD or FasterRCNN, which we carefully modify and customly train to satisfy the challenging requirements of our customers. Our models are specifically optimized to provide a very high accuracy and, at the same time, allow for sufficient refresh rates on current embedded systems, such as NVidia Jetson Modules or the latest FPGAs.

Detection of Cars and other Objects

Image Super-Resolution

Image super resolution is an actively researched topic in Computer Vision. Image Super-resolution is the process of enhancing the quality of the image by increasing its resolution using deep learning methods.

When Image-super resolution is paired with object detection, the number of objects detected in the super resolved images is more than the number of objects detected in the native resolution. It is also very useful in increasing the quality of the dataset.

AMDC is working with a number of state-of-the-art super resolution algorithms to produce high quality images for various applications.

Close

Synthetic Training Data for Computer Vision

With the increase in ML applications, there is also an increase in the demand for quality data to train ML models. Oftentimes there is no or only limited training data readily available. In such cases, synthetic training data plays a vital role to train ML models. AMDC is capable of generating synthetic data for various computer vision applications such as classification and object detection models.

Synthetic Data for Object Detection and Classification

AMDC has built a pipeline that makes use of an open source graphical software to place 3D models into a scene and to render training images. The scene consists of several parameters that are modified each time when an image is rendered, thus simulating various real world scenarios such as sunlight, shadow, dust particles etc. The rendered images are saved along with information required for the training (ground truth information). The resulting dataset is directly used for training the AI models.

Synthetic Image

Training Data Generation Service

Object Detection is the process of localizing and classifying certain objects within an image.

AMDC has developed a system to create 3D scenes of environments and place objects therein. Images of the objects can be rendered from any perspective and any distance with different terrain, weather, textures, and daylight conditions. Corresponding labels can be generated as well.

Based on a 3D-scene, more than 50.000 training images with labels can be generated in one day (24h) by a standard high performance PC. The system is also capable to provide data in the IR-spectrum.

Based on customer requirements we can provide a large amount of labeled images in short time. It is possible to mix the images with labeled photos to improve the robustness of the trained AI-system. If you have any questions, please contact info@amdc.de .

Synthetic Images

Close

Defence

AMDC was founded as a consultancy company for Air and Missile Defence, which is still a core business area. We are convinced that well equipped armed forces are essential for every nation to secure peace and our intention is to contribute to this objective. In the defence branch we focus on advising the German Bundeswehr and allied armed forces in the matters of military research and technology. Key activities lie in the analysis, conception, and evaluation of modern and modular Air Defence and Missile Systems, using innovative technologies. Our staff has many years of experience in the areas of missiles, sensors, command & control, communication, and threat analysis. In addition, through cooperation with the defence industry, universities, and scientific aerospace research institutions, we apply highly specialized know-how to our analyses. Due to these close collaborations, we keep our knowledge up-to-date and make the most recent research results available to our customers. AMDC conducts system analyses under various aspects. Usually, the primary focus is on the performance or effectiveness of the missile system. It is determined not only by the missile itself, but also by the other system components, in particular, the separate surveillance and fire control sensors. Also, safety, cost-effectiveness, and environmental compatibility are important aspects of system analyses. The most valuable tool for system analyses is the simulation.

Learn More

Computational Engineering

In our Computational Engineering Department we deal with the development and application of simulation models to solve the complex physical problems we encounter throughout the work in our main branches Aerospace, Machine Learning and Defence.

Our Computational Engineering Services are also available to external customers who require tailor-made and scientifically reliable simulation solutions in a timely manner.

Learn More