Studienprojekte

Projekte sind ein essenzieller Bestandteil des Curriculums von Mobile Computing. Die Studierenden bekommen die Möglichkeit, das im Zuge ihres Studiums erworbene theoretische Wissen selbst praktisch umzusetzen. Ein sowohl für Studierende als auch für Lehrende immer wieder spannendes Unterrichtskonzept, in dem schon erfolgreiche Startups wie z.B. runtastic und Butleroy ihre Anfänge gefunden haben.

Universal Data Labelling Tool (UDLT)

Zeitraum
Mar 2019 - Jul 2019
FH Studierende
Samuel Stadler
FH BetreuerIn
Alexander Palmanshofer BSc MSc

Ziel

Universal Data Labeling Tool, or short UDLT, is a web-based application that supports the process of labelling audio and video files. Machine learning grows on popularity and the data to label increases year by year. These circumstances require good tools to create a well-defined ground truth for training accurate machine learning models.

Umsetzung

UDLT organizes the data to be labeled in projects. Each of these projects are either of the type video or audio. A single video project can contain up to 4 video files and additional sensor data, like the temperature, pressure or an audio track. The data can then be labeled frame by frame and within one frame multiple sensor data can also be tagged at once. The ability to label several data at once speeds up the process of creating an accurate dataset.

An audio project decodes the specified audio file and presents it as a waveform and spectrogram. Both representations are zoomable so that more detailed information can be displayed. UDLT also integrates a media player so the audio files can be played and paused within the tool.

In order to be able to actually tag the data of a video or audio project, UDLT makes label tracks available. A label track may be described as an audio track, but instead of audio information, it includes various captions that classify the actual data within a time range. Sometimes it is required to overlap labels, this is the case when a single frame of a video contains multiple information that can all be useful for training a neural network. UDLT solves this requirement by adding multiple label tracks to keep them clean of intersections. In further consequence the labeled data can be exported as a JSON file to use it for training machine learning models.

All projects including their files, tags and labels get stored on the server and can be shared with colleagues.