Hyun

The link you referred has the solution. You are saying you want to do? Do what you want mate!

We are trying to make the downloading process be encapsulated as "tfds" package does. But 'tfds' gives *.tfrecord file. We are looking for other ways to get image-extension too.

The output for the confidence values, which mean that there is an object or not, are allocated to each bounding boxes. So, Your input label shaped (BATCH_SIZE, 1) is confusing...

The labels are differed by datasets and how you define the anchor boxes. I didn't make this repo to fit in a specific dataset. If you propose a dataset for...

@HangJie720 Hello, I used datasets that I make on my own. This code is kind of template to utilize for various datasets. Thank you for asking!

@Jaren1907 Basically, BlazeFace uses Anchor box. Hence, the label form should follow Anchor scheme. To make anchor boxes, I recommend you to refer RetinaNet repo (https://github.com/fizyr/keras-retinanet)

Thanks for opening an issue. From any dataset, I have to make a processing logic for it. There have been many questions about the label form for actual training. I...