How can i use the validation set in object detection using deep learning?

I use faster-RCNN using pre-trained model(ResNet-50) for object detection.

(https://kr.mathworks.com/help/vision/examples/object-detection-using-faster-r-cnn-deep-learning.html)

I'm fine-tunning the last layers using a custom dataset.

As far as I know, when you have a small dataset k-fold cross validation is the right technique for evaluation.

But, this example does not seem to go through a validation procedure. Is there something I'm missing?

Answer 1

[Divya Gaddipati]

The trainFasterRCNNObjectDetector does not support the “ValidationData”, “ValidationFrequency” and “ValidationPatience” options as of now. Hence, it is not possible to directly feed the validation data to the trainingOptions parameter.

Refer to this documentation for more information on how to use trainFasterRCNNObjectDetector function and the Tips section at the end of the page:

https://www.mathworks.com/help/vision/ref/trainfasterrcnnobjectdetector.html

Possible workaround for the k-fold validation could be to run a loop “k” times.

• In each loop, split the train set randomly into two sets, where one set will be used for training and the other for validation. You can use crossvalind or cvpartition functions for this purpose.
• Use this new train set for training the model and after the training is done, use the validation set to evaluate the model.

This above step is done “k” times.

For more information of k-fold cross validation, you can refer to the following links:

1) https://www.mathworks.com/help/bioinfo/ref/crossvalind.html

2) https://www.mathworks.com/help/stats/cvpartition.html