The strong emphasis on classification experiments is motivated by their facilitation of rapid content descriptor development. The data-sets are carefully selected to be specialized tools in this process. Despite their different focusses, all sub-sets are non-overlapping, thus can be combined to the \textit{MVD-MIX} data-set, which is intended for similarity retrieval and recommendation experiments.
Subset | Classes | Videos | Artists |
---|---|---|---|
MVD-VIS | 8 | 800 | 490 |
Music genre classification with visual features | |||
MVD-MM | 8 | 800 | 550 |
Multi-modal music genre classification | |||
MVD-MIX | 16 | 1600 | 1040 |
Extended multi-modal music genre classification |
The dataset creation was preceded by the selection of genres. For the MVD-VIS dataset eight orthogonal classes with minimum overlap were defined. This aim was accomplished by restricting the search on clearly defined sub-genres. For the MVD-MM dataset eight top-level genres with high inter-genre overlaps were selected. Additional avoidance of overlaps between the genres of these two subsets allow for a combination into the bigger MVD-MIX dataset. Each genre consists of 100 selected videos. Entries for the MVD-VIS and MVD-MM datasets were selected primarily by their audible properties. This decision was based on the introducing definition of Music Video Information Retrieval - a cross-domain approach to MIR problems.