search

endernewton / tf-faster-rcnn

Tensorflow Faster RCNN for Object Detection
https://arxiv.org/pdf/1702.02138.pdf
MIT License
3.65k stars 1.58k forks source link

About resnet block3 #402

Open qcyay opened 5 years ago

qcyay commented 5 years ago

In tf-faster-rcnn-master\lib\nets\resnet_v1.py,the 140th line in the code is self._blocks = [resnet_v1_block('block1', base_depth=64, num_units=3, stride=2), resnet_v1_block('block2', base_depth=128, num_units=4, stride=2),

use stride 1 for the last conv4 layer

                  resnet_v1_block('block3', base_depth=256, num_units=23, stride=1),
                  resnet_v1_block('block4', base_depth=512, num_units=3, stride=1)]

But why the stride of the block3 is 1,not 2?Because in the paper the output size becomes 7 7 from 14 14 after passing the block3.

xytjcxy commented 5 years ago

Because the feat_stride equals to 16 rather than 32,so it has to be 1. However,if I change the stride from 1 to 2 and feat_stride from 16 to 32 , the result will be worse! And no matter what the stride is ,the dimensionality in fc will be 1000!

qcyay commented 5 years ago

Thanks for your explanation!