Best practices for image annotation

Hi All,

I am working to label a bunch of images for object detection. Specifically I am looking at images from cameras in sewer pipes. One specific challenge I am looking at is labelling roots. Roots come in several different varieties - clumped as a ball or many different strands. I’m looking for some guidance on how to draw bounding boxes.

Any guidance is much appreciated.

Raw Image:
Many boxes:
One box:

Example of large mass of roots:

It comes down to what you are trying to ascertain. Provide annotations in the way you want output annotations to look, not the way you think the network might want them to look.

From the examples you provide it looks more like an image classification problem to me. Why do you need BBs?

I built a system originally with just classification and got great results on validation set but poor results on similar looking test data. I eventually would need bounding boxes anyway so I moved to an object detection system which seems to be working better than my original classification system on test data.


You need to first identify which task you need the dataset for. If you are someone looking forward to creating your own dataset for your various computer vision tasks then Labellerr is a platform where you can perform the annotation part. Labellerr is an Automated AI and Data Annotation SAAS platform.

Here is the link to the platform:(

Link to the blogs: (

First, identify which task you want to create a dataset for and then get started. The task can be like image segmentation, object detection, image classification. Here is a video that demonstrates image annotation.