Confused about implement of vlad core

[gordon636798]

why line 25. code is `v = tf.expand_dims(inputs, -1) + C' in stead of subtraction

[Leputa]

It confuse me too. and I can't understand line 11 why use_bias = False...

[YiZhu89]

It confuse me too. and I can't understand line 11 why use_bias = False...

"There are two versions of the NetVLAD layer in the code: layerVLAD and layerVLADv2 (they are very similar - do the diff). layerVLADv2 follows the paper exactly, while layerVLAD is a version where we fix $b_k$ parameters (i.e. they are not updated during training). Assuming that input descriptors are L2-normalized (don't use layerVLADotherwise!!), which seems to anyway work better with VLAD and NetVLAD when combined with conv5 descriptors, the assignment into clusters can be done more simply by scalar product (i.e. initially $w_k$=L2-normalized cluster centre, $b_k$=0). This version seems to converge faster, so it is the version we use in all our experiments. It might well be that layerVLADv2 can get you better results - it should work at least as well as layerVLAD, unless you are overfitting too much, as it has the capacity to do everything layerVLAD can do, and more!"

https://github.com/Relja/netvlad ReadMe_more.md