search

DetectionTeamUCAS / Cascade-RCNN_Tensorflow

MIT License
89 stars 34 forks source link

mobilenet V2 #15

Open huaxv opened 4 years ago

huaxv commented 4 years ago

I used mobilenet V2 with CasCade R-CNN network training, but compared to Resnet101, the accuracy rate is reduced by 20%, but the paper of mobilenet V2 will not decrease so much, which is about 2%. The number of times is okay. What causes this?

huaxv commented 4 years ago

I see in the original version of Mobilenet V2 that there is a compression factor, but I did not find this coefficient in the code. My accuracy rate has dropped too much, will it be related to this coefficient?