Difference between revisions of "2014:Query by Tapping"

From MIREX Wiki
m (Overview)
m (Step 0: Indexing the MIDIs collection)
Line 60: Line 60:
 
  ...
 
  ...
  
Output indexed files are placed into <dir_workspace_root>.
+
Output indexed files are placed into <code><dir_workspace_root></code>.
  
 
=== Step 1: Training ===
 
=== Step 1: Training ===

Revision as of 21:05, 27 August 2014

Overview

The main purpose of QBT (Query by Tapping) is to evaluate MIR system in retrieving ground-truth MIDI files by tapping the onset of music notes to the microphone. This task provides query files in wave format as well as the corresponding human-label onset time in symbolic format. For this year's QBT task, we have three corpora for evaluation:

  • Roger Jang's MIR-QBT: This dataset contains both wav files (recorded via microphone) and onset files (human-labeled onset time).
    • 890 onset & .wav queries; 136 ground-truth MIDI files
  • Show Hsiao's QBT_symbolic: This dataset contains only onset files (obtained from the user's tapping on keyboard).
    • 410 onset queries; 143 ground-truth MIDI files (128 of which have at least one query)
  • Kaneshiro et al.'s QBT-Extended: This dataset contains only onset files (obtained from users tapping on a touchscreen). Documentation can be found here.
    • 3,365 onset queries (1,412 from long-term memory and 1,953 from short-term memory) from 60 participants; 51 ground-truth MIDI files
    • A hidden dataset is currently being collected, from 20 new participants

Discussions for 2014

CCRMA is very excited to be hosting the QBT task this year!

Submission deadline is September 16.

Any questions or suggestions can be added directly here, or you can send us an email, qbt | at | ccrma dot stanford dot edu

New for 2014

  • Added QBT-Extended dataset
  • New subtask combining rhythm and pitch contours (QBT-Extended only)
  • Minor modifications to command format - moving more toward a training/test paradigm, accommodates hidden test datasets
  • Please make sure your list of candidates is ranked, as we will be assigning variable points for Top-10, 5, and 1 matches.

Task description

  • Evaluations are performed separately on each dataset

Subtask 1: QBT with symbolic input

  • Test database: The set of ground-truth MIDI files corresponding to each dataset.
  • Query files: Text files of onset times to retrieve target MIDIs from all datasets listed above. These onset files can help participant concentrate on similarity matching instead of onset detection. Onset files derived from .wav files cannot guarantee to have perfect detection result from original wav query files.
  • Evaluation: Return top 10 candidates for each query file. 1 point is scored for a hit in the top 10 and 0 is scored otherwise (Top-10 hit rate). We may also consider Top-5 and Top-1 scoring.

Subtask 2: QBT with wave input

  • Test database: About 150 ground-truth monophonic MIDI files in MIR-QBT.
  • Query files: About 800 wave files of tapping recordings to retrieve MIDIs in MIR-QBT.
  • Evaluation: Return top 10 candidates for each query file. 1 point is scored for a hit in the top 10 and 0 is scored otherwise (Top-10 hit rate). We may also consider Top-5 and Top-1 scoring.

Subtask 3: QBT-Extended with symbolic input (new for 2014)

  • This subtask uses a longer query vector concatenating tap times and (pitch) positions.
  • Development dataset: The set of ground-truth MIDI files in the QBT-Extended dataset. Both onset times and MIDI note numbers are used.
  • Query files: Text files of onset times in the QBT-Extended dataset (long-term and short-term memory queries). Both onset times and vertical coordinates of tasks are considered.
  • Development evaluation: Return top 10 candidates for each query file in the development dataset. 1 point is scored for a hit in the top 10 and 0 is scored otherwise (Top-10 hit rate). We may also consider Top-5 and Top-1 scoring.
  • Test evaluation: Return top 10 candidates for each query file in the hidden dataset. 1 point is scored for a hit in the top 10 and 0 is scored otherwise (Top-10 hit rate). We may also consider Top-5 and Top-1 scoring.

Command formats

Step 0: Indexing the MIDIs collection

If your algorithm needs to pre-process (e.g., index) the database, your code should do so using the following command-line format (Note that this step is not required unless you want to index or preprocess the MIDI database).

Command format should look like this:

indexing <dbMidi.list> <dir_workspace_root>

where <dbMidi.list> is the input list of database midi files named as uniq_key.mid. For example:

QBT/database/00001.mid
QBT/database/00002.mid
QBT/database/00003.mid
QBT/database/00004.mid
...

Output indexed files are placed into <dir_workspace_root>.

Step 1: Training

The command format should be like this:

qbtTraining <dbMidi_list> <query_file_list> [dir_workspace_root]

Where <dbMidi_list> is a list of the MIDI files in the database to match against (see Step 0), and <query_file_list> maps each query to its associated ground truth. You can use [dir_workspace_root] to store any temporary indexing/database structures. (You can omit [dir_workspace_root] if you do not need it at all.)

Per-task input specification

If the input query files are onset files (for subtask 1), then the format of <query_file_list> is like this (tab-separated):

qbtQuery/query_00001.onset	00001.mid
qbtQuery/query_00002.onset	00001.mid
qbtQuery/query_00003.onset	00002.mid
...

See details of Onset files format

If the input query files are wave files (for subtask 2), the the format of <query_file_list> is like this (tab-separated):

qbtQuery/query_00001.wav	00001.mid
qbtQuery/query_00002.wav	00001.mid
qbtQuery/query_00003.wav	00002.mid
...

If the input query files are .onset and .y_onset files (for subtask 3), then the format of <query_file_list> is like this (tab-separated):

qbtQuery/query_00001.onset	qbtQuery/query_00001.y_onset	00001.mid
qbtQuery/query_00002.onset	qbtQuery/query_00002.y_onset	00001.mid
qbtQuery/query_00003.onset	qbtQuery/query_00003.y_onset	00002.mid
...

See details of Onset files format


Onset files format

To preserve compatibility with the original task, the QBT-E query files share the same .onset file extension as previous symbolic input query datasets.

An .onset file is a space-separated text file of elapsed onset times (in milliseconds) from the first onset, which is always 0.0. Example of 5 onsets:

0.0 479.922 720.069 976.071 1215.694

For subtask 3, the additional dimension (position/pitch/contour) is provided in a file with the same name, but with extension .y_onset. Similarly to an .onset file, .y_onset files are space-separated text files, the same length as the corresponding .onset file, that lists the absolute vertical position of each tap on the touchscreen. Example (corresponding to above):

291.000000 293.500000 305.500000 302.000000 239.000000

Step 2: Testing

The command format should be like this:

qbtTesting <query_file_list> <resultFile> [dir_workspace_root]

Where <query_file_list> are input queries, and <resultFile> is the filename where your script should store results. You can use [dir_workspace_root] to store any temporary indexing/database structures. (You can omit [dir_workspace_root] if you do not need it at all.)

<resultFile> gives ranked top-10 candidates for each query (note that ranking of the candidates is new for 2014). For instance <resultFile> should have the following format for subtasks 1 and 3:

qbtQuery/query_00001.onset: 00025 01003 02200 ... 
qbtQuery/query_00002.onset: 01547 02313 07653 ... 
qbtQuery/query_00003.onset: 03142 00320 00973 ... 
...

And for subtask 2:

qbtQuery/query_00001.wav: 00025 01003 02200 ... 
qbtQuery/query_00002.wav: 01547 02313 07653 ... 
qbtQuery/query_00003.wav: 03142 00320 00973 ... 
...

Where 00025 is the top-ranked (closest match) MIDI file for query_00001, followed by 01003, 02200, etc. Note that the output should be the names of the MIDI files (e.g., 00025 means 00025.mid); they are not necessary 5-digit numbers.

Potential Participants

Jorge Herrera, Hyung-Suk Kim, and Blair Kaneshiro, CCRMA, qbt at ccrma dot stanford dot edu

References

Chen JCC, and Chen ALP (1998). Query by rhythm: An approach for song retrieval in music databases. Research Issues in Data Engineering, Proceedings of IEEE Eighth International Workshop on Continuous-Media Databases and Applications, 139-146.

Eisenberg G, Batke JM, and Sikora T (2004). BeatBank - an MPEG-7 compliant query by tapping system. Audio Engineering Society Convention 116, paper 6136.

Eisenberg G, Batke JM, and Sikora T (2004). Efficiently computable similarity measures for query by tapping systems. Proceedings of the Seventh International Conference on Digital Audio Effects (DAFx'04), Naples, Italy, 189-192.

Hanna P, and Robine M (2009) Query by tapping system based on alignment algorithm. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1881-1884.

Hébert S, and Peretz I (1997). Recognition of music in long-term memory: Are melodic and temporal patterns equal partners? Memory & Cognition 25:4, 518-533.

Jang JSR, Lee HR, and Yeh CH (2001). Query by tapping: A new paradigm for content-based music retrieval from acoustic input. Advances in Multimedia Information Processing PCM, 590-597.

Kaneshiro B, Kim HS, Herrera J, Oh J, Berger J, and Slaney M (2013). QBT-extended: An annotated dataset of melodically contoured tapped queries. Proceedings of the 14th International Society for Music Information Retrieval Conference, Curitiba, Brazil, 329-334.

Peters G, Anthony C, and Schwartz M (2005). Song search and retrieval by tapping. Proceedings of the National Conference on Artificial Intelligence 20, 1696.

Peters G, Cukierman D, Anthony C, and Schwartz M (2006). Online music search by tapping. Ambient Intelligence in Everyday Life, 178-197.