I am working on an object detection model and have thought of looking into stratified splits for the dataset.

Now since I am doing object detection I have a variable number of "labels" for every image because in each image there is variable number of occurrences for each object I am looking for (car, truck, motorbike, etc.).

Obviously single-label stratification does not apply.

From what I understand multi-label stratification is only applicable if there are basically label "features" that we know are always present, which does not seem the case here.

My question is... is there a way to perform stratified split in this case so that in each split there is roughly the same number of cars/trucks/bikes/etc.? (Or is it going to actually improve the results at all?)

  • $\begingroup$ group them on the basis of the feature of the multilabel $\endgroup$ – Rithik Banerjee Jul 30 '20 at 14:20
  • $\begingroup$ @RithikBanerjee can you please explain in more detail? The very definition of multilabel seems to be that there are multiple features which can be considered labels. $\endgroup$ – Rares Dima Jul 31 '20 at 5:04
  • $\begingroup$ @RaresDima Did you find an optimal solution? $\endgroup$ – Sadra Naddaf Feb 18 at 23:05
  • $\begingroup$ @SadraNaddaf no. There isn't one. The best you can do is consider this as a partitioning problem and approach it in a greedy way because of the massive amount of data. $\endgroup$ – Rares Dima Mar 11 at 16:08
  • $\begingroup$ @sadranaddaf If you explain and clarify your question I can post my experience/code as an answer here. If you have roughly split each label to train/dev/test with 70%/10%/15% so then you have ended up with a different number of images in each set but they roughly have an equal % of each class label. Is it the same as what you want? I did a similar thing on my OD task. $\endgroup$ – Sadra Naddaf Mar 12 at 3:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.