Which paramaters can improve detector performance?

Hello everyone, I trained Faster RCNN and SSD object detectors with my own dataset with 5 classes… My results are: Faster RCNN average precision 64% and SSD average precision 45% … I was inspirated by examples on matworks page… which parameters should I change or increase to get better results? I mean if increasing minibatchsize would give me better average precision or what should i change ? Thanks for answers… I mean which parameters in training options can improve my results?

Answer 1

[Sachin]

Hi @Adrian Kleffler

I understand that you want to improve the detector performance.

Following suggestions might be helpful to you:

1. Increase mini-batch size : Increasing the mini batch size can speed up training and reduce the variance of the model.
2. Data augmentation : By using Data augmentation techniques like flipping, rotation, resizing can help you to improve the performance of your model.
3. Adjust learning-rate – Experiment with learning rate might be helpful.
4. Adjust hyper-parameters : You can consider changing the hyperparameters such as number of epochs, number of layers in your mode, optimization algorithm .

Refer the following MATLAB Documentation page :

https://in.mathworks.com/help/vision/ug/object-detection-using-faster-r-cnn-deep-learning.html

Thanks

Sachin