Publishing a trained Inception-ResNet-V2 model, as reported for the ISMIR paper.
Run standalone_inference_over_image.py with the trained model to use it on your images.
Update 16.05.2019:
Added models that are also significantly smaller than the originally trained model:
| Feature Extractor | Validation Performance (COCO-AP in %) |
|---|---|
| MobileNet | 62.67 |
| ResNet50 | 78.14 |
| Inception-ResNet-V2 (24.04.2019) | 80.04 |
| Inception-ResNet-V2 (16.05.2019) | 80.97 |