search

chuanqi305 / MobileNetv2-SSDLite

Caffe implementation of SSD and SSDLite detection on MobileNetv2, converted from tensorflow.
MIT License
448 stars 231 forks source link

result box has offset for ssdlite_mobilenet_v2_coco net #34

Closed tianylijun closed 6 years ago

tianylijun commented 6 years ago

i convert ssdlite_mobilenet_v2_coco_2018_05_09 model from tensorflow to caffe, the result boxs get from caffe (I use your ssd caffe with relu6 version) all have some offset, any suggest about the issuse?

ssd-3 ssd-2 ssd-1

I get there is has_tf_pad in base_conv_layer.cpp, how to add this param into prototxt?????

chuanqi305 commented 6 years ago

@tianylijun Make sure use ReLU6 to replace ReLU

tianylijun commented 6 years ago

@tianylijun Make sure use ReLU6 to replace ReLU

@chuanqi305

offset problem is the pad diff between TensorFlow and caffe， I fix it add tf_pad =SAME in caffe prototxt（hope you update the gen_model.py with tf_pad =SAME support, thanks)， and I replace relu with relu6. But the result still has some diff with TensorFlow， and the dog recog as bird。 ssd

tianylijun commented 6 years ago

I have fix it

runauto commented 5 years ago

@tianylijun how to fix it ?i meet the same error as you.