Relabel image in object detection project based on seperate project (single image classification confusion matrix results)

Community Help
, ,
1

Hi,

Setup

My setup is a object detection project (that categorizes defects) and a separate single label classification that takes the isolated bounding boxes and classifies the defects. When looking at the confusion matrix for one of the single label classification models and the specific images that are misclassified I notice obvious bad annotations that I want to relabel in the object detection project (original dataset):

image
image1596×1093 134 KB

Problem

The problem is the apply tags to images button, that would help me locate the images and relabel only applies the tags to my single label image classification project, which is just subset copy of the object detection project. I want to apply those tags to the object detection project, such that I can relabel the original images instead.

Solutions?
Any way to easily solve this? Am I missing a feature that already solves this problem?

  • I think being able to train classification models in the same project as my object detection projects could solve the problem.

  • Otherwise the option to sync tags across projects of different types would be really useful.

2

Hello! Thank you for your message. In Roboflow, this particular grouping of annotations across projects is called an Annotation Group, and it is configured when you first create a project as described in our documentation here:

For more information about annotation groups, please refer to the blog post here: What is an annotation group?

1 Like
3

Hi - thank you for the quick answer!

It looks like this might be helpful… However, will it be able to sync the annotations across projects (object detection & single label image classification)? In my object detection project I labelled bounding boxes with the following tag-structure:
category-type1-type2-class. E.g. a bbox tag could be annotated cat1-F-B-3

Is this a suboptimal setup? Should I restructure my projects? Is it better to only tag one thing per project?

For now I’m just interested in training an object detection model to find the bounding box of different categories and then isolating that bounding box and classifying it, using a single image classification model.
But given a image in the object detection project might have 5 bounding boxes, can I still make a copy and sync it with a single label project? I just modify the classes when generating versions…

Hopefully that made some sense. Hoping I can benefit further from your expertise:)

4

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.