2010:Audio Music Similarity and Retrieval Results

From MIREX Wiki

Introduction

These are the results for the 2010 running of the Audio Music Similarity and Retrieval task set. For background information about this task set please refer to the Audio Music Similarity and Retrieval page.

Each system was given 7000 songs chosen from IMIRSEL's "uspop", "uscrap" and "american" "classical" and "sundry" collections. Each system then returned a 7000x7000 distance matrix. 100 songs were randomly selected from the 10 genre groups (10 per genre) as queries and the first 5 most highly ranked songs out of the 7000 were extracted for each query (after filtering out the query itself, returned results from the same artist were also omitted). Then, for each query, the returned results (candidates) from all participants were grouped and were evaluated by human graders using the Evalutron 6000 grading system. Each individual query/candidate set was evaluated by a single grader. For each query/candidate pair, graders provided two scores. Graders were asked to provide 1 categorical BROAD score with 3 categories: NS,SS,VS as explained below, and one FINE score (in the range from 0 to 100). A description and analysis is provided below.

The systems read in 30 second audio clips as their raw data. The same 30 second clips were used in the grading stage.


General Legend

Team ID

Sub code Submission name Abstract Contributors
BWL1 MTG-AMS PDF Dmitry Bogdanov, Joan Serrà, Nicolas Wack, Perfecto Herrera
PS1 PS09 PDF Tim Pohle, Dominik Schnitzer
PSS1 PSS10 PDF Tim Pohle, Klaus Seyerlehner, Dominik Schnitzer
RZ1 RND PDF Rainer Zufall
SSPK2 cbmr_sim PDF Klaus Seyerlehner, Markus Schedl, Tim Pohle, Peter Knees
TLN1 MarsyasSimilarity PDF George Tzanetakis, Steven Ness, Mathieu Lagrange
TLN2 Post-Processing 1 of Marsyas similarity results PDF George Tzanetakis, Mathieu Lagrange, Steven Ness
TLN3 Post-Processing 2 of Marsyas similarity results PDF George Tzanetakis, Mathieu Lagrange, Steven Ness

Broad Categories

NS = Not Similar
SS = Somewhat Similar
VS = Very Similar

Understanding Summary Measures

Fine = Has a range from 0 (failure) to 100 (perfection).
Broad = Has a range from 0 (failure) to 2 (perfection) as each query/candidate pair is scored with either NS=0, SS=1 or VS=2.

Human Evaluation

Overall Summary Results

<csv p=3>2010/ams/AMS2010summary_evalutron.csv</csv>
Note:RZ1 is the random result for comparing purpose.

Friedman's Tests

Friedman's Test (FINE Scores)

The Friedman test was run in MATLAB against the Fine summary data over the 100 queries.
Command: [c,m,h,gnames] = multcompare(stats, 'ctype', 'tukey-kramer','estimate', 'friedman', 'alpha', 0.05);

<csv p=3>2010/ams/evalutron.fine.friedman.tukeyKramerHSD.csv</csv>

2010AMS.evalutron.fine.friedman.tukeyKramerHSD.png

Friedman's Test (BROAD Scores)

The Friedman test was run in MATLAB against the BROAD summary data over the 100 queries.
Command: [c,m,h,gnames] = multcompare(stats, 'ctype', 'tukey-kramer','estimate', 'friedman', 'alpha', 0.05);

<csv p=3>2010/ams/evalutron.cat.friedman.tukeyKramerHSD.csv</csv>

2010AMS.evalutron.cat.friedman.tukeyKramerHSD.png


Summary Results by Query

FINE Scores

These are the mean FINE scores per query assigned by Evalutron graders. The FINE scores for the 5 candidates returned per algorithm, per query, have been averaged. Values are bounded between 0 and 100. A perfect score would be 100. Genre labels have been included for reference.

<csv p=1>2010/ams/fine_scores.csv</csv>

BROAD Scores

These are the mean BROAD scores per query assigned by Evalutron graders. The BROAD scores for the 5 candidates returned per algorithm, per query, have been averaged. Values are bounded between 0 (not similar) and 2 (very similar). A perfect score would be 2. Genre labels have been included for reference.

<csv p=1>2010/ams/cat_scores.csv</csv>

Raw Scores

The raw data derived from the Evalutron 6000 human evaluations are located on the 2010:Audio Music Similarity and Retrieval Raw Data page.

Metadata and Distance Space Evaluation

The following reports provide evaluation statistics based on analysis of the distance space and metadata matches and include:

  • Neighbourhood clustering by artist, album and genre
  • Artist-filtered genre clustering
  • How often the triangular inequality holds
  • Statistics on 'hubs' (tracks similar to many tracks) and orphans (tracks that are not similar to any other tracks at N results).

Reports

BWL1 = Dmitry Bogdanov, Joan Serrà, Nicolas Wack, Perfecto Herrera
PS1 = Tim Pohle, Dominik Schnitzer
PSS1 = Tim Pohle, Klaus Seyerlehner, Dominik Schnitzer
RZ1 = Rainer Zufall
SSPK2 = Klaus Seyerlehner, Markus Schedl, Tim Pohle, Peter Knees
TLN1 = George Tzanetakis, Mathieu Lagrange, Steven Ness
TLN2 = George Tzanetakis, Mathieu Lagrange, Steven Ness
TLN3 = George Tzanetakis, Mathieu Lagrange, Steven Ness

Run Times

<csv>2010/ams/audiosim.runtime.csv</csv>