1
$\begingroup$

Dear Research Community,

I am new to object detection research and currently working on a project which will detect people using YOLO from a top-down camera image.

While looking through tutorials for YOLO, I see all tutorials asks us to download 'pre-trained weights', but to the best of my knowledge, none of the tutorials talks about why(or how) we are using the pre-trained weights and whether we can train the YOLO model without it.

I will appreciate it if someone can shed some light into what are the pre-trained weights in YOLO?

With best, Azmyin

Improve this question
$\endgroup$

2 Answers 2

Reset to default
1
$\begingroup$

From what I understand, CNN's like Yolo here, require loads of images to train. These images must have lots of variations in them to be robust enough (in your case variability in types of people, lighting conditions, viewing angles, etc.). Therefore, a popular use case is to use transfer learning where a pre-trained model is used and the user can then train a new model by starting from this point. The idea here, I believe, is that the first few layers have already learnt what kind of features to extract, what to look for in the images, etc. It just needs to learn the last (or maybe last few) layers to work for your specific use case.

Another reason to use pre-trained models is to be able to use readily available images and get comfortable with the framework and see it in practice before starting your own work.

Intel's openvino has a lot of pre-trained models here including one for pedestrian (person) detection which can be converted and used with yolo. This could be used as a starting point in your work!

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks Mr. Kurup. I agree, I was able to train a Darknet-53 model to a custom dataset with help of a pre-trained model. Trying to do that without the pre-trained weights caused the training time to become prohibitively long $\endgroup$
    – Azmyin Md. Kamal
    Dec 23, 2020 at 1:53
0
$\begingroup$

I found this post, which sheds some light to this question. https://machinelearningspace.com/yolov3-tensorflow-2-part-3/

Hope this helps someone out

Improve this answer
$\endgroup$

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.