2
$\begingroup$

I want to use an object detection model for some use case.

I started with the YOLOv3, because I need to be able to perform detection on multiple images in less than few seconds (Even if it is not for real-time / video).

So, I successfully loaded and generated predictions using a keras model with pretrained weights.

The problem is that the pre-trained weights for this model have been generated with the COCO dataset, which contains very few classes (80).

What really surprises me is that all the pre-trained weights I can found for this type of algorithms use the COCO dataset, and none of them use the Open Images Dataset V4 (which contains 600 classes).

Is there a reason for that ? And if there is not, is there a place where I can find such pre-trained weights ?

$\endgroup$
1
$\begingroup$

Darknet has a yolo implementation using Open Images check this page (scroll down up to the end, just before the citation)

You can simply use wget https://pjreddie.com/media/files/yolov3-openimages.weights

$\endgroup$
4
  • $\begingroup$ I tried to use these weights, but it look like there is only few classes. When I tried to do a detection with a turtle image, it is pretty confident that it is a bird (like with the weights trained with COCO, which have not been trained to recognize turtles), and with a pizza picture it give no prediction at all (with the weights trained with COCO the prediction is good). $\endgroup$
    – Nakeuh
    Apr 16 '19 at 15:59
  • $\begingroup$ Try retraining the last layer $\endgroup$ Apr 16 '19 at 16:00
  • $\begingroup$ How to pass labels to train last layer? model has to predict boundary boxes right? $\endgroup$
    – LUZO
    Jul 5 '19 at 14:31
  • $\begingroup$ learnopencv.com/… check this tutorial $\endgroup$ Jul 5 '19 at 15:06

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.