Open Source Tool for Data Annotation

Project description

The goal of this project is to develop an open source tool, modified and improved based on the dataset capabilities which can detect an object and 3D bound them. In the end this model should work with a prediction model that will decrease the time and computing power used for video inputs for frame-by-frame interpretation.

Context

HERE Technologies is a multinational group dealing with mapping, location data and related automotive services to individuals and companies. HERE is interested in improving their data science tooling, more specifically their data annotation tooling for Lidar points datasets.

Results

The end result, besides the experimenting notebooks with the models presented above, is a notebook focused entirely on our second approach using YOLOv5. We call this notebook the “Multi Tool” as we intended to have steps from data importing and preparation, YOLO being applied and then going to Lidar interpretation using PointNet, but this last step we could not achieve.

During our second approach we used the original repository for YOLOV5 and we managed to feed the KITTI data into it to be analyzed. We experimented with the different weight versions YOLO has and the best results were of course provided by the largest one, yolov5x. The impressive part was that the model was able to run on a normal computer, without using the dedicated graphics card and got the results quite fast and accurate.

Methodology

Research will be required to answer the research questions that have been established. A research framework, specifically the "Five Strategies Framework for Research," will be used to manage and define the meaning of this research. For answering research questions, this framework employs five strategies. These are the following: Field, Library, Workshop, Lab, and Showroom. As shown in the figure, the strategies Library and Showroom are placed in the available work domain, whereas Field and Lab are placed in the Application domain, and Workshop is placed centrally because it is applicable to both.

About the project group

Our project group is formed with two ICT students, Vlad being from the Technology stream while Wing comes from Software. Hence, our attention for the project was drifted more on the programming side of things. The project took a total of 18 weeks where we researched, tested and developed a multitool for our client in the use of Automotive technologies.