search

chenzhi1992 / TensorRT-SSD

Use TensorRT API to implement Caffe-SSD， SSD（channel pruning）， Mobilenet-SSD
251 stars 84 forks source link

The confidence of the prediction will exceed 1 after addding extra prior box #33

Open 408550969 opened 6 years ago

408550969 commented 6 years ago

I changed the number of categories(30) and the proportion of boxes(add 3 aspect ratio in the 4_3 layer,and 4 in the fc7 and conv6_2 layer). I changed all the 21 to 30 in pluginImplement.cpp and pluginImplement.h. And changed the appropriate number is in cudaSoftmax. When I predict, the precision is very low. And the confidence level is greater than 1. What are the reasons and what else do I need to modify?

chenzhi1992 commented 6 years ago

You should modify the aspect ratio parameter of PriorBox Layer in pluginImplement.cpp

408550969 commented 6 years ago

Yes, I changed it.

408550969 commented 6 years ago

I find that it is not a question of prior box, but a question of category. I used SSD to train my own dataset, and the results were not correct.