I use DSSD to detect the football. The image sets are labeled by myself. I train SSD mode and DSSD mode on those image sets. I run
python examples/ssd/ssd_pascal_resnet_321.py
to train SSD mode. The parameter which i adjust is the minibatch size to be enough for GPU memory. And the result is 0.73.
I think that the result could be better if I adjust the hyper parameters of the SSD such as scales and aspect ratios for default boxes, the mean vaule of the image and so on. And I try to change aspect ratio, but the result is less than 0.2 that is quite bad.
So I want to ask how to change the hyper parameters to get better reslut.
I use DSSD to detect the football. The image sets are labeled by myself. I train SSD mode and DSSD mode on those image sets. I run
python examples/ssd/ssd_pascal_resnet_321.py
to train SSD mode. The parameter which i adjust is the minibatch size to be enough for GPU memory. And the result is 0.73.
I think that the result could be better if I adjust the hyper parameters of the SSD such as scales and aspect ratios for default boxes, the mean vaule of the image and so on. And I try to change aspect ratio, but the result is less than 0.2 that is quite bad.
So I want to ask how to change the hyper parameters to get better reslut.