flowVision.dev()

In ColorInMotion the technical and economical feasibility for a software product for colorizing black-and-white footage could be successfully shown. With the insight gained in this project a cooperation with HS-Art was established to develop a market-ready software application for colorization. The main goal of the application is to improve the speed and quality of the colorization compared to existing processes. This will be achieved through utilization of state-of-the art AI-based algorithms that assists operators without limiting their artistic freedom.

The automotive division of the NUMTEC brand of Alpine Metal Tech produces machines for inspection and machinining of aluminium wheels. Several of these machines need to detect the valve hole to correctly orientate the wheel. In this project the robustness and the integration of this valve hole detection process was improved in several machines to reduce the error rate, speed up the cycle time, and to improve the usability.

Through practical use of the placement algorithm at customer sights valuable insight into the behaviour of the algorithm in production could be gathered. Based on these insights the placement the fill rate of the algorithm and its runtime were improved.

At the moment colorization of black and white movies and film material is a very time consuming and costly job. This project showed that with a specialized software that uses state of the art algorithms and models it is possible to increase the efficiency of the process so that the required manual work can be significantly reduced which reduces time and cost of a colorization project.

To show this I developed a small application in which different algorithms and models can be tested. In the project state of the art models for different computer vision problems, e.g. Video Object Segmentation, Optical Flow, or Automatic Colorization were analysed, adapted, integrated, and combined to create an efficient workflow for the colorization process. An important task in the project is also to find out how the user interface should be designed so that the models can be used intuitively. Additionally new traditional algorithms were developed to also support the workflow. Some examples of these developed algorithms are an automatic index mask refinement and an edge-aware brush.

The goal of this project was to speed up the Pick-it-Easy Robot of KNAPP. The computation of the k-nearest-neighbors (kNN) of points in a 3D point cloud is an elementary computation for many other algorithms used in the system. Because of that a CUDA implementation of the kNN search was developed. The implementation utilizes the special structure in which the 3D data are stored to be able to yield fast results. To be able to easily use the kNN search inside the Pick-it-Easy Robot it was integrated with the framework of the Point Cloud Library.

The DIAMANT-Film restoration software offers many filters to repair different kinds of film defects. Lately the demand of restoring video material has increased. As the defects that occur when digitizing video tapes are very different from film defects new filters needed to be developed that specifically address these issues.

One of the main tasks of KNAPP's Pick-it-Easy Robot is to place items into a target container. For this an internal representation of the current state of the container had to be developed to find an optimal placement position. Based on research papers on the Online Bin Packing Problem and the Container Loading Problem I developed a data structure and an algorithm which suggests a list of possible positions sorted by their "optimality". The algorithm can handle real world problems like bridging small gaps between items, overhanging items, and restrictions due to the robot. For different uses cases it offers several strategies for finding and sorting the placement positions.