[MIREX] MIREX评测介绍

MIREX作为国际最权威音频检索评测大赛，竟然在百度上找不到任何介绍，只有几个与什么搜狗、腾讯获得什么成绩相关的检索内容，相比而言，TRECVID的内容收到重视多了...由于研究生阶段主要研究音频领域，需要对整个领域有一个大致的了解，感觉还是从MIREX入手比较合适，所以借此机会也与大家分享一记。

MIREX全称Music Information Retrieval Evaluation eXchange，即音乐信息检索评测，至于eXchange放在这不太清楚什么意思，或许与“交流”类似的含义吧，比赛由IMIRSEL承办，每个子项目由任务组织者设计并管理，这些任务组织者基本就是各个领域的领头专家。

【最普适的任务：音频分类任务】

  • Audio Classification (Train/Test) Tasks

包含了以下几个子任务：1. 美国流行音乐、拉丁音乐、韩国流行音乐的流派分类，2. 音乐情感分类、韩国流行音乐情感分类，3. 古典音乐的作曲家鉴别。这个任务做了很多年，感觉准确率到达一个瓶颈，不同任务的准确率基本上就稳定在0.65~0.8之间。

【音频相似度和检索】

  • Audio Music Similarity and Retrieval

音频相似度和检索，7000首30s的歌曲，返回一个稀疏矩阵，对每首歌返回相似度前100名的歌曲及相似度。看看应用场景吧 A music similarity system can help a music consumer find new music by finding the music that is most musically similar to specific query songs (or is nearest to songs that the consumer already likes). 其实不太清楚这种相似性度量是通过哪个衡量标准：节拍、速度、调式、节奏、旋律、和声、和弦，中的一个还是几个。

 

【符号旋律相似性】

  • Symbolic Melodic Similarity

计算旋律相似性，应该指的是通过MIDI的旋律符号，比较旋律的相似性。Retrieve the most similar items from a collection of symbolic pieces, given a symbolic query, and rank them by melodic similarity. There will be only 1 task this year which comprises a set of six "base" monophonic MIDI queries to be matched against a monophonic MIDI collection. 类似于以下结构信息

ALTDEU
CUT[Das Hildebrandslied]
REG[Europa, Mitteleuropa, Deutschland]
KEY[A0001  04  G 4/2]
MEL[1_  3b_3b_4_4_  5__5__
    0_5__5_  5_6_7b_5_  5__0_
    5_  5_6_7b_5_  6b__5__
    0_5_4_3b_  5_3b_3b__
    0_3b_3b_3b_  4_4_5__  5__0_
    5_  4_3b_3b_3b_  2__1__
    0_5_5_.4  3b__0_
    5_  6b_5_5_3b_  4__5__
    0_4_3b3b1_  1_-6_-7__  1__. //] >>
FCT[Romanze, Ballade, Lied]
Format

 

【结构分段】

  • Structural Segmentation

The segment structure (or form) is one of the most important musical parameters. It is furthermore special because musical structure -- especially in popular music genres (e.g. verse, chorus, etc.) -- is accessible to everybody: it needs no particular musical knowledge. 输入一段音乐，输出的是对这段音乐的分段信息，如以下格式

0.000    5.223    A
5.223    15.101   B
15.101   20.334   A
Format

 

【多基频检测与跟踪】

  • Multiple Fundamental Frequency Estimation & Tracking

Estimation，将每一固定10ms内的基频检测出来；Tracking，将基频的持续长度检测出来。感觉类似于对象检测与跟踪啊，检测与跟踪一般都相辅相成的，所以算法应该是互相交叉的。所以

Example :
time	F01	F02	F03	
time	F01	F02	F03	F04
time	...	...	...	...
which might look like:
0.78	146.83	220.00	349.23
0.79	349.23	146.83	369.99	220.00	
0.80	...	...	...	...

For the second task, for each row, the file should contain the onset, offset and the F0 of each note event separated by a tab, ordered in terms of onset times:
onset	offset F01
onset	offset F02
...	... ...
which might look like:
0.68	1.20	349.23
0.72	1.02	220.00
...	...	...
Format

 

【音频节奏检测】

  • Audio Tempo Estimation

Submitted programs should output two tempi (a slower tempo, T1, and a faster tempo, T2) as well as the strength of T1 relative to T2 (0-1). The relative strength ST2 (not output) is simply 1 - ST1. The tempo estimates from each algorithm should be written to a text file in the following format

T1<tab>T2<tab>ST1
E.g.
60	180	0.7

评价标准是

P = ST1 * TT1 + (1 - ST1) * TT2

where ST1 is the relative perceptual strength of T1 (given by groundtruth data, varies from 0 to 1.0), TT1 is the ability of the algorithm to identify T1 to within 8%, and TT2 is the ability of the algorithm to identify T2 to within 8%. No credit will be given for tempi other than T1 and T2. 然后奇怪的事情就在这，这里说ST1是given by groudtruth data，那么自己预测的ST1不参与评测吗？

The algorithm with the best average P-score will achieve the highest rank in the task.

【音频标签分类】

  • Audio Tag Classification

与Traing/Test任务类似，不同的是这里允许一个样本对应多个不同标签，所以最后的输出是一个稀疏矩阵，如下形式

I.e.:
 <example path and filename>\t<tag classification>\t<affinity>\n
E.g.:
 /data/file1.wav    rock      0.9
 /data/file1.wav    guitar    0.7
 /data/file1.wav    vocal     0.3
 /data/file2.wav    rock      0.5
 ...
Format

 

【歌单识别】

  • Set List Identification 

应用场景：演唱会。可以分解为两个子任务，即歌曲检测与跟踪

1. To identify the order of songs which be performed in a live concert.

In this sub task, the participants known the the artist and artist's studio song collection. Assigning a live concert audio and studio songs collection of a specific artist, all songs in live concert are included in studio songs collection, to identify the order of songs in this live concert.

2. To identify the start/end time of each song in song sequence

In this sub task, the participants known the artist, artist's studio song collection and the song sequence. Assigning a live concert audio, song sequence and studio songs collection of a specific artist, all songs in live concert are included in studio songs collection, to identify start time and end time of each song in the live concert.

这两个任务是衔接的子任务，都是给定的歌曲列表：子任务一的输出是这些歌曲在演唱会中的顺序；子任务二的输出是上述排出序的歌曲在演唱会中分别的起始终止时间。

 

【】

  • Audio Onset Detection

【】

  • Audio Offset Detection

【】

  • Audio Beat Tracking

【】

  • Audio Key Detection

【】

  • Audio Downbeat Detection

【】

  • Real-time Audio to Score Alignment(a.k.a Score Following)

【音频翻唱歌曲识别】

  • Audio Cover Song Identification

翻唱歌曲识别，比歌曲相似度任务更难，据我所知主要与旋律相关，要求的输出格式如下，一个完全矩阵

Example distance matrix 0.1
1    /path/to/audio/file/track1.wav
2    /path/to/audio/file/track2.wav
3    /path/to/audio/file/track3.wav
4    /path/to/audio/file/track4.wav
5    /path/to/audio/file/track5.wav
Q/R   1        2        3        4        5
1     0.00000  1.24100  0.2e-4   0.42559  0.21313
3     50.2e-4  0.62640  0.00000  0.38000  0.15152
Format

评价标准如下

The following evaluation metrics will be computed for each submission: 1. Total number of covers identified in top 10；2. Mean number of covers identified in top 10 (average performance)；3. Mean (arithmetic) of Avg. Precisions；4. Mean rank of first correctly identified cover。话说1和2是一个意思吧，MAP在10时的值；3是平均准确率，应该还跟内部位置有关；4是第一个识别正确的cover song的排名

 

【重复主题章节的发现】

  • Discovery of Repeated Themes & Sections

Algorithms that take a single piece of music as input, and output a list of patterns repeated within that piece. Also known as intra-opus discovery. 输入：一段音乐；输出：在这段音乐里重复出现的模式。那么所谓的模式是什么呢？For the purposes of this task, a pattern is defined as a set of ontime-pitch pairs that occurs at least twice (i.e., is repeated at least once) in a piece of music. The second, third, etc. occurrences of the pattern will likely be shifted in time and perhaps also transposed, relative to the first occurrence. Ideally an algorithm would be able to discover all exact and inexact occurrences of a pattern within a piece, so in evaluating this task we are interested in both (1) whether an algorithm can discover one occurrence, up to time shift and transposition, and (2) to what extent it can find all occurrences. It has been pointed out by Lartillot and Toiviainen (2007) among others that as well as ontime-pitch patterns, there are various types of repeating pattern (e.g., ontimes alone, duration, contour, harmony, etc.). For the sake of simplicity, the current task is restricted to ontime-pitch pairs.

【】

  • Audio Melody Extraction

【】

  • Query by Singing/Humming

【】

  • Audio Chord Estimation

【】

  • Singing Voice Separation

【】

  • Audio Fingerprinting

 

posted @ 2016-03-29 00:11  CarlGoodman  阅读(2181)  评论(0)  编辑  收藏  举报