I want to annotate custom objects in around 4.000 images, where each image contains many objects. I cannot accomplish the work by hand as you can understand. I searched on google and here on stackoverflow, but the solutions are based on "common" annotations let's say, such as car, horse, person, house, etc. I want to annotate custom datasets that they do not exist as "common"/"ready" in the platforms. How can I proceed?

I need to have polygons' labeling and not just rectangles on each automatically annotated object. And have the annotations in .json format. Any ideas?

I want to annotate custom objects in around 4.000 images, where each image contains many objects. I cannot accomplish the work by hand as you can understand. I searched on google and here on stackoverflow, but the solutions are based on "common" annotations let's say, such as car, horse, person, house, etc. I want to annotate custom datasets that they do not exist as "common"/"ready" in the platforms. How can I proceed?

I need to have polygons' labeling and not just rectangles on each automatically annotated object. And have the annotations in .json format. Any ideas?

• annotations
• conv-neural-network

1 Answer 1

I can suggest the following strategy to tackle the annotation task you want to complete:

1. Use open source annotation platform like cvat to annotate small batch first (500 images for example) - cvat

2. Train a detection model on the annotated images -> run inference on a second batch of images (let's say another 500) -> revisit model predictions. It should take less time to revisit predictions

3. Repeat 1 and 2 until you finish the data that you need to annotate.

4. Regarding the polygon requirement, I recommend to use a zero-shot segmentation model like SAM where you prompt it with ground truth rectangles that you get from the annotation phase suggested above.