sh1r0 / caffe-android-demo

An android caffe demo app exploiting caffe pre-trained ImageNet model for image classification
https://github.com/sh1r0/caffe-android-lib
MIT License
255 stars 165 forks source link

The performance of libcaffe.so/libcaffe_jni.so of armeabi-v7a is 3 times slower than that of arm64-v8a #59

Closed hgffly closed 7 years ago

hgffly commented 7 years ago

Hi sh1r0, I have tried this program with my trained model. For recognizing an image, it's take about 1.5s. But due to the requirement of integrating caffe with my developed app, I need to use libcaffe.so/libcaffe_jni.so of armeabi-v7a. I built this version of libraries from your another project caffe-android-lib. To evaluate the performance, I replace caffe.so/caffe_jni.so of arm64-v8a with the ones I built. I just found that it takes about 4.8s to recognize an image, about 3 times slower. Is it correctly? Do I miss something? How should I correct or improve it ? Thanks