ccmusic-database/Guzheng_Tech99

Intro

For the 99 recordings, silence is first removed, which is done based on the annotation, targeting the parts where there is no technique annotation. Then all recordings are uniformly segmented into fixed-length segments of 3 seconds. After segmentation, clips shorter than 3 seconds are zero padded. This padding approach, unlike circular padding, is adopted specifically for frame-level detection tasks to prevent the introduction of extraneous information. Regarding the dataset split, since the dataset consists of 99 recordings, we split it at the recording level. The data is partitioned into training, validation, and testing subsets in a 79:10:10 ratio, roughly 8:1:1.

Demo

https://huggingface.co/spaces/ccmusic-database/Guzheng_Tech99

Usage

from modelscope import snapshot_download
model_dir = snapshot_download("ccmusic-database/Guzheng_Tech99")

Maintenance

git clone [email protected]:ccmusic-database/Guzheng_Tech99
cd Guzheng_Tech99

Results

Backbone	Mel	CQT	Chroma
ViT-B-16	0.705	0.518	0.508
Swin-T	0.849	0.783	0.766
VGG19	0.862	0.799	0.665
EfficientNet-V2-L	0.783	0.812	0.697
ConvNeXt-B	0.849	0.849	0.805
ResNet101	0.638	0.830	0.707
SqueezeNet1.1	0.831	0.814	0.780
Average	0.788	0.772	0.704

Dataset

https://huggingface.co/datasets/ccmusic-database/Guzheng_Tech99

Mirror

https://www.modelscope.cn/models/ccmusic-database/Guzheng_Tech99

Evaluation

https://github.com/monetjoe/ccmusic_eval/tree/tech99

Cite

@dataset{zhaorui_liu_2021_5676893,
  author       = {Monan Zhou, Shenyang Xu, Zhaorui Liu, Zhaowen Wang, Feng Yu, Wei Li and Baoqiang Han},
  title        = {CCMusic: an Open and Diverse Database for Chinese Music Information Retrieval Research},
  month        = {mar},
  year         = {2024},
  publisher    = {HuggingFace},
  version      = {1.2},
  url          = {https://huggingface.co/ccmusic-database}
}