thoughts on the annotation process + new addition to dl-annotator: ability to drag and drop files

[radekosmulski]

I annotated a bunch of Mourning Dove files using dl-annotator. I have two observations:

1. I underestimated how good the train data is - at least for the Mourning Dove, the calls in the recordings are frequent, sampling randomly from the file has a little chance ending up with a 5 second segment without a Mourning Dove call. On the other hand, I also underestimated how many birds are there in the background, bird calls of different species. This leads me to believing that just presenting a model with a 5 second segment and asking it to look for what it might use for identification is not ideal. The obvious solution are bounding boxes... but I don't feel good about this. It requires implementing a model that is not straightforward at all and the annotation process is so taxing... I have some thoughts about an architecture that would be between an SSD and regular CNN that maybe could work off simplified labels...
2. Working with multiple short files poses a challenge - when I go to open another file through the system dialog for opening files, it always brings me to the top of the directory. It takes forever to find the next file in line. I could hack together some GUI component to help with this (maybe read files in succession in a directory? that would work great) but for now I am thinking that just having an ability to drag files over should suffice - this way I can have an explorer window open and hopefully be able to identify easier which file is next 🤔 Implemented ability to drag and drop files in hopes that this assumption is correct.