[question]I have changed the network from ssd+vgg to ssd+resnet, do you know how to generate prior_boxes_ssd300.pkl?

[KevinYuk]

As title.

[fferroni]

If you have changed the initial convolutional stacks from a VGG to a residual type architecture, you simply need to ensure that the output convolved tensors are the same after each of these two stacks. If you do so, you needn't change anything in the prior boxes.

[KevinYuk]

@fferroni Thanks very much for your comments. As is known, the topology of VGG and resnet are different from each other, and their output convolution tensors are not the same. My understanding is that prior_boxes_ssd300.pkl is the weights of VGG network part in SSD. That means if we change the imagenet part of the SSD from VGG to resnet, we have to pre-train this new imagenet network(aka: resnet) in the new SSD topology. Am I right?

[gcucurull]

@KevinYuk prior_boxes_ssd300.pkl are the default bounding boxes used by SSD, not the weights of the VGG part.

[KevinYuk]

Hi @gcucurull ,

Thanks for your comments. If prior_boxes_ssd300.pkl are the default bounding boxes used by SSD, could I set prior_boxes_ssd300.pkl any value if I want to train the SSD network?

Thanks a lot.