Researchers Juan F. Montesinos, Olga Slizovskaia, and Gloria Haro from Universitat Pompeu Fabra have introduced Solos, a new dataset for evaluating source separation algorithms in music information retrieval.

The Solos dataset comprises 755 recordings spanning 13 instrument categories, sourced from YouTube, with an average duration of 5:16 minutes per recording.

Notably, 8 out of the 13 instrument categories feature HD resolution recordings.

Each recording in the dataset includes OpenPose skeletons for body and hand movements, with timestamps for useful segments.

OpenPose utilizes neural networks to predict skeletons through the analysis of confidence maps and part affinity fields.

The dataset and additional information are available on the researchers' website, aimed at assisting researchers in training deep neural networks for various audio-visual tasks.

Solos contributes to the advancement of multimodal techniques in the field of music information retrieval.