RedenFlow help
Bringing in a dataset you already have
Most of what you read about RedenFlow assumes you are starting from a folder of photographs and no labels. Plenty of people are not. If you already have a labelled dataset, or an hour of video and no stills at all, you do not have to start again.
Three ways in
COCO JSON
The format most annotation tools export. Boxes and polygon segmentation both come across.
YOLO TXT
A folder of images and a folder of .txt files. Boxes and polygons both, again.
A video
Frames pulled out at an interval you choose, so you get a spread rather than a hundred near-identical stills.
You do not have to say which format it is
Point the import wizard at the folder. It works out the structure on its own: whether this is COCO or YOLO, where the images live, where the labels live, and how many of each there are. It shows you what it found and waits, so a folder that turns out to hold three images rather than three thousand is something you see before importing rather than afterwards.
Class names come across with their names and get a colour each. Renaming or deleting a class afterwards rewrites every annotation file in the project in one go, so the labels never drift out of step with the class list.
Polygons survive
This is the detail worth checking before you import anything anywhere. A tool that reads only bounding boxes will quietly flatten a segmentation dataset into rectangles, and the work that went into the outlines is gone.
Both importers read polygon segmentation as polygons. If your dataset has outlines, you keep outlines, and you can train a segmentation model on them.
Questions people ask
Which formats can it read?
COCO JSON and YOLO TXT. Both are read with boxes and with polygon segmentation, so a segmentation dataset comes in as polygons rather than being flattened to rectangles.
Do I have to tell it which one I have?
No. You point it at the folder and it works out the structure, including where the images are and where the labels are, and shows you what it found before it imports anything.
What about Pascal VOC XML?
Not supported. COCO and YOLO cover most of what people arrive with, and a converter to one of those is usually a single command in whatever tool produced the VOC files.
I have video, not photographs.
Import the video and RedenFlow pulls frames out of it, taking every Nth frame so you get a spread rather than a hundred near-identical stills from one second. Video import is a Pro feature.
Will the class names come across?
Yes, with their names, and each gets a colour. You can rename or delete a class afterwards and every annotation file in the project is rewritten to match, in one go.
Can I mix an imported dataset with photographs I label myself?
Yes. Once it is imported it is an ordinary project. Add more pictures, label them by hand or with one click, and train on the lot.
Have a format that is not on this list? Tell us what it is. It is useful to know what people actually arrive with.