Caffe: what's the VGG_ILSVRC_16_layers_fc_reduced.caffemodel ?

[Sundrops]

Hi Wei Liu,

Thanks for sharing the code. Is models_VGGNet_VOC0712_SSD_300x300.caffemode based on VGG_ILSVRC_16_layers_fc_reduced.caffemodel by VOC2007 and VOC2012 dataset ? Thanks a lot

[weiliu89]

I did net surgery on original VGG16 net by converting fc6 and fc7 from fully connected layer to convolutional layer, and subsample the parameters from them, following what DeepLab did. All released SSD models are built on top of the reduced VGGNet.

[shawn-tian]

@weiliu89 the filters are down sampled from 4096 to 1024, is it randomly selected? The kernel size of fc6 is changed from 6_6 to 3_3, how did you down sample the filters here? Thank you

[Sundrops]

@weiliu89 thanks a lot !

[zxDeepDiver]

@weiliu89 @shawn-tian I also get into this problem when converting other models into fully convolutional networks. Could you tell us how you down sample the filters? Thank you.