2
$\begingroup$

Seeking clarity on single class object detection model using ML. I have prepared a custom database for this purpose up to 400 images which is split in 80%-20% as training and testing data-set. These are top view only images. The data collection followed the basic guidelines provided at here.

enter image description here

The objective now is to detect the zebra crossing in below contexts.

enter image description here

Model is failing in terms of accuracy. Although it identifies the class but fails to localize correctly in some contexts(greater extent).The desirable result is in blue whereas the model throws up red. enter image description here

What changes are required to training dataset to rectify this?

EDIT1:- Tensorflow object detection API is used for this task. Detection accuracy attained is above 90%. Looking forward for suggestions to fix Localisation issue.

EDIT2:- The illustrations here are only for outlining the issue.All my images are real world pics.

$\endgroup$
2
$\begingroup$

I have the following suggestion:

  • The size of your data-set is small, you should increase the sample size. You also have to use data augmentation methods for improving the performance, translation, orientation and blurring are highly recommended.

  • For your task I guess the most popular approach is what which is discussed in YOLO paper, and it is highly recommended using that if you want to have object detection not just localization and recognition for good performance.

  • I also suggest you providing real data for your task. The images of first row are artificially made. You have to provide real data which your classifier is going to face. The distribution of your test time should be like validation and train time if you want to have good performance.

| improve this answer | |
$\endgroup$
  • $\begingroup$ It worked with first and third recommendations. :-) $\endgroup$ – addcolor Mar 1 '18 at 15:03
  • $\begingroup$ @addcolor I happy it helped you. $\endgroup$ – Media Mar 2 '18 at 13:29

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.