Closed mikekestemont closed 4 years ago
Since Nikolay has issues in running the proper versions of TensorFlow I'm taking over the classification part, we will fine-tune the ECCV models on the instruments collection. Code is already ready, just have to extract the crops based on cytomine.
It seems that the script works with tensorflow_gpu-1.14.0. So, I will run experiments on Monday (hopefully)
That's supercool! Over here we have good results for the detection of instruments as well (check the overleaf paper for a pre-view) so if Mike helps with writing a nice story on why musical instruments are interesting we'll have the paper :-)
Perfect. I've read quite a number of papers over the last days and will start writing tomorrow morning.
There is now a notebook for the saliency-maps visualization and filter-visualization in ./saliency_maps/
We can use it on the models which have been trained for the classification task on the images of the testing-set.
Superb! I will check it out ASAP.
Start with simple classification model: both for large and full, and for individual instruments and instrument categories.