search

ChunML / ssd-tf2

A super clean implementation of SSD (Single Shot MultiBox Detector) made possible by Tensorflow 2.0
MIT License
115 stars 54 forks source link

There is only one BN in network, is that useful? #9

Closed wang-tf closed 4 years ago

wang-tf commented 4 years ago

In my opinion, there is no information about batch normalization in paper. And there is only first feature layer used this operation in code. Is that useful ? Can we use it for every feature layer before compute_heads ?

https://github.com/ChunML/ssd-tf2/blob/53e481ade7016c6d83e008a84f7cd18e59c75242/network.py#L92

ChunML commented 4 years ago

In the paper, the authors stated:

Since, as pointed out in [12], conv4 3 has a different feature scale compared to the other layers, we use the L2 normalization technique introduced in [12] to scale the feature norm at each location in the feature map to 20 and learn the scale during back propagation.

Because there weren't any equivalent implementations for Tensorflow, I just used a BatchNorm layer instead. I tried to implement L2 normalization layer on my own, but the loss didn't converge as expected.