CODEWITHSUNDEEP
image-processingcomputer-visionobject-detection
Mostapha Aloui
Mostapha Aloui

Reputation: 23

Is same-location pictures good enough to train an object detection model?

I'm currently working on a custom object detection for which I need to gather data (labeled images of the object). The object is rare enough that I only managed to find around 50 distinct pictures of it on google images, but there are a couple of two-minutes videos that I downloaded and saved every frame of them as a distinct picture(which amount to over 9000 (not a meme) images). Is this a good approach or not ?

Upvotes: 0

Views: 129

Answers (1)

Renat Gilmanov
Renat Gilmanov

Reputation: 17895

It is definitely a good approach, especially in case of rare objects. There are not so many other ways to create high-quality training set. Check additional details here:

Another great example demonstrates you can get great results with basic equipment and a tiny piece of creativity:

enter image description here

(see the video)

You just have to make sure:

  • videos are different to avoid overfitting
  • it is not a problem to annotate huge amount of images
  • more data can be added in case results are not that good
  • you can automate at least some steps

I've been busy developing training set generation solution for quite some time (see some examples) and see no other way but to create a fully or semi automated solution based on video.

enter image description here

(see the complete set)

Upvotes: 1

Related Questions