Hello, I am using image.manual recipe to label images. I am using prodigy v1.8.3.
Here are the steps that I am doing -
- Move the images I want to label in a single folder.
- Create a new dataset with command prodigy dataset trying_set "Trying out prodigy" --author Me
- Use the image.manual recipe by command prodigy image.manual trying_set dataset/cars --label CAR
- Then by opening http://localhost:8080 in my browser, I see my images. I mark the bounding boxes on image, click on the green tick(accept), then label next image and so on. Till now, I can go back and edit any annotations using the history.
- In the end, I see No Tasks Available written as I labeled all images in the directory. I click save button on the upper left part of the screen, and close my browser, and exit the process in the terminal.
Now, if I want to iterate through the images and see what I have annotated, how do I do it?
I tried using prodigy image.manual trying_set dataset/cars --label CAR but it gives me the screen No Tasks Available and an empty history. Is it possible to go through all images and edit bounding boxes?
Please let me know if any more information is required.
Hi! The dataset argument you’re passing in on the command line is always the name of the set to save the annotations to – but you can export the data from an existing dataset and then load it in to correct the data or add more annotations. Prodigy’s input and output formats are the same, so the exported JSONL files are also valid input files.
For example, you can export the data like this:
prodigy db-out trying_set > annotations.jsonl
And then re-annotate it like this:
prodigy image.manual trying_set_corrected annotations.jsonl --loader jsonl --label CAR
--loader jsonl, you’re telling Prodigy to treat the input source as a JSONL file instead of a directory of images (which is the default).
I’d also recommend using a new dataset to store the corrected annotations (in my example, I called it
trying_set_corrected). This keeps your datasets clean and tidy and makes it easy to start over if you make a mistake (without messing up your original data).
Thanks, got it.
So, I have to finish annotating, save,export then reload from the exported jsonl.
Is it possible to edit the bounding box i.e. move the vertices of the bounding box? That would make annotation quite easy, sometimes we need to be precise in annotations.
Yes, exactly. One of the principles of Prodigy datasets is that they're append-only and that individual records are immutable. This lets you always reconstruct a given annotation and what the annotator saw and labelled.
Not at the moment, no! The manual image annotation UI is still fairly new and we first wanted to focus on the basics and quick/rough annotation features and see how precise we need to make the annotation workflow to cover the different use cases.
From your experience, what's more common: wrong sizing (e.g. box is too small or too large) or wrong positioning? Or both?
Actually, both issues are quite common.
Yes, please do add this feature soon!!