Is the feat in spp_detect.m the final feature used in fully connected layer?

ShaoqingRen / SPP_net

SPP_net : Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

364 stars 237 forks source link

Is the feat in spp_detect.m the final feature used in fully connected layer? #22

Closed prettyzjx closed 9 years ago

prettyzjx commented 9 years ago

RT In spp_detect.m line 40 to line 43:

feat = spp_features_convX(im, spm_im_size, [], use_gpu);
feat = spp_features_convX_to_poolX(spp_model.spp_pooler, feat, boxes, false);
feat = spp_poolX_to_fcX(feat, spp_model.training_opts.layer, spp_model, use_gpu);
feat = spp_scale_features(feat, spp_model.training_opts.feat_norm_mean);

Is the feat the final feature used in fully connected layer? In a word, can I use the feat as the extracted feature in other ways?

ShaoqingRen commented 9 years ago

Hi prettyzjx,

yes, feat is the feature after fully connected layer. It can be used in other ways, but whether should be finetuned and from which layer should be validated.

prettyzjx commented 9 years ago

Hi,

One more question please. What is the candidate region? I cannot find it on the paper. Thanks!

ShaoqingRen commented 9 years ago

@prettyzjx

The candidate regions are generated by Windows Proposal methods (e.g., Selective Search and Edge Box)