I have followed provided notebook for dataset splitting with one notable change - instead of downloading the dataset I have imported my own using importer.Import... . I have tested it in a couple of ways:
a. Importing COCO and exporting YOLO
b. Importing YOLO, and exporting YOLO
In each case, I have tested both types of splitters.
- Stratified split seems to not work without some value for validation split
- They never correctly split the data for both of them - in terms of pure % wise division, the smaller the val or (val & test) part where the poorer the split was. For example, 80,10,10 split for Stratified split resulted in 92,4,4 samples per split.
Is that because it prioritizes equal class split (I have two) over the number of samples?
As for the negative samples in COCO format - importer throws an error if I try to load a bunch of images with empty annotations - meaning no objects of interest are in those images.
It works for YOLO .txt empty labels file though
I have followed provided notebook for dataset splitting with one notable change - instead of downloading the dataset I have imported my own using importer.Import... . I have tested it in a couple of ways:
a. Importing COCO and exporting YOLO
b. Importing YOLO, and exporting YOLO
In each case, I have tested both types of splitters.
Is that because it prioritizes equal class split (I have two) over the number of samples?
As for the negative samples in COCO format - importer throws an error if I try to load a bunch of images with empty annotations - meaning no objects of interest are in those images.
It works for YOLO .txt empty labels file though