ISMIR Paper Explorer (2015)

ISMIR Paper ExplorerClick image to open the demo website in a new window
(
Please note: The survey is not enabled for this demo.)

This is joint work together with Thomas Low and Christian Hentschel who are supervised by Andreas Nürnberger and Harald Sack, respectively.

Challenge

Explore the cumulative proceedings of the annual conference of the International Society for Music Information Retrieval (ISMIR).

Solution

Apply Visual Berrypicking, i.e. aligning local neighborhood maps to create the impression of panning a large (global) map.

Core Technologies

  • visualization and interaction techniques: Visual Berrypicking
  • backend implementation: Python
  • frontend implementation: HTML5 (jQuery)

Visual Berrypicking (2014)


This is joint work together with Thomas Low and Christian Hentschel who are supervised by Andreas Nürnberger and Harald Sack, respectively. The demo video was created by Thomas Low.

Challenge

With increasing collection size and complexity of the similarity metric, projection accuracy rapidly degrades and computational costs prevent online map generation.

Solution

Create the impression of panning a large (global) map by aligning inexpensive small maps showing local neighborhoods.

Core Technologies

  • local map generation: multidimensional scaling (MDS)
  • map alignment: Procrustes analysis
  • backend implementation: Python, Java
  • frontend implementation: HTML5 (jQuery)

Beatles History Explorer (2013)

beatles history explorerClick image to open the demo website in a new window
(This public demo does not allow playing the music tracks and only displays pixelated cover images to avoid copyright issues.)

Challenge

Music collections usually grow over time. How can this change be appropriately visualized?

Solution

Change collection map only as little as possible to incorporate new tracks and cross-fade between consecutive map versions using animations.

Core Technologies

  • map generation and incremental updates: (landmark) multidimensional scaling (MDS) combined with Procrustes analysis, incremental growing self-organizing map (GSOM), stochastic neighbor embedding (SNE), neighbor retrieval visualizer (NeRV)
  • backend implementation: Java, Python
  • frontend implementation: HTML5 (jQuery)

BLE-X Navigator (2013)

EFB Ble-XThe BLE-X Navigator search interface can be accessed at http://ble-x.de/explorer (in German).

Challenge

Develop an exploratory web search interface for the research reports and conference proceedings published by the European Research Association for Sheet Metal Working (Europäische Forschungsgesellschaft für Blechverarbeitung e.V.)

Core Technologies

  • OCR analysis of scanned documents with custom document parser for structure extraction and support of within-document search
  • search result visualization as classical list (left) and similarity space (center) using multidimensional scaling (MDS) and Adaptive SpringLens
  • backend implementation: Java, Python
  • frontend implementation: HTML5 (jQuery)

MusicGalaxy (2010)

Challenge

How to deal with unavoidable projection errors (“wormholes”) in map-based visualizations?

Solution

Visualize wormholes through adaptive distortion of the map and facilitate navigation “through wormholes” into distant regions.

Core Technologies

  • map generation: landmark multidimensional scaling (Landmark MDS)
  • map distortion: Adaptive SpringLens technique (grid-based distortion)
  • track distance/similarity measure: manually adaptable
  • implementation: Java using QT Jambi for visualization

Beatles Explorer (2008)

Challenge

Different users with a different musical background may have different ideas about what makes two music tracks sound similar. Hence, a similarity-based structuring of a music collection should reflect personal views.

Solution

Learn a personalized distance measure from the way a user interacts with the collection.

Core Technologies

  • hexagonal Growing Self-Organizing Map (GSOM) for structuring the collection into groups of similar tracks
  • user feedback by assigning tracks to different groups (drag & drop)
  • automatic adaptation of the track distance measure based on the new group assignment through quadratic programming
  • implementation: Java using the processing.org framework for visualization