I have a data set of 25 images. I wish to run Faster RCNN or yolov3 object detection models on this images.I want to create my custom trained model and get weights after running say 10 epochs. Later I can save these weights and use that for prediction. Build a model, train on my image data set and get weights. Is it possible?
$\begingroup$
$\endgroup$
2
-
1$\begingroup$ You want to train an object detection model on 25 images? How would that work? $\endgroup$– theblackipsOct 3, 2019 at 11:18
-
$\begingroup$ Since any object detection network like yolo or faster rcnn would take more time to train on many images in a cpu, I want to build yolo or faster rcnn network architecture, train on only 20 images(with 2 classes, 2 kinds of flowers) and validate on 5 images with few epochs.Randomly initialize weights from layer 1 and not to use pre trained model and weights derived from training with imagenet or cats and dogs data set. This exercise is for learning and understanding the architecture. $\endgroup$– MalathiOct 3, 2019 at 16:23
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Maybe you could first train the model on a large sized flower dataset, that way the conv layer will be optimised to detect flowers and then do transfer learning on that model using your custom dataset freezing all but the last layer. I am not sure though, 25 images seem way too less. Try data augmentation on those images like horizontal and vertical flips, noise, shear etc.
