718c41634f
1.项目后端整体迁移至PaddleOCR-NCNN算法，已通过基本的兼容性测试 2.工程改为使用CMake组织，后续为了更好地兼容第三方库，不再提供QMake工程 3.重整权利声明文件，重整代码工程，确保最小化侵权风险 Log: 切换后端至PaddleOCR-NCNN，切换工程为CMake Change-Id: I4d5d2c5d37505a4a24b389b1a4c5d12f17bfa38c
3.2 KiB
caffe-android-lib https://github.com/sh1r0/caffe-android-lib
mini-caffe https://github.com/luoyetx/mini-caffe
openblas-0.2.20 https://github.com/xianyi/OpenBLAS
ncnn https://github.com/Tencent/ncnn
squeezenet_v1.1 https://github.com/DeepScale/SqueezeNet/tree/master/SqueezeNet_v1.1
mobilenet_v1 https://github.com/shicai/MobileNet-Caffe
vgg16 https://gist.github.com/ksimonyan/211839e770f7b538e2d8
Host platform and compiler configuration:
fedora 27, android-ndk-r15c, target arch = arm64-v8a
we manually update openblas package to version 0.2.20 in caffe-android-lib for better performance
Device: Nexus 6p
OS: LineageOS 15.1(Android 8.1.0), ROM newly flashed without any third-party APP installed
CPU: Snapdragon 810 (Cortex-A57 2.0GHz x 4 + Cortex-A53 1.55GHz x 4)
RAM: 3G
Benchmark method:
Run squeezenet, mobilenet inference 23 times in a loop, discard the first three warmup records, and then calculate the average inference time
Run vgg169 times in a loop, discard the first warmup record, and then calculate the average inference time
Since the system may force SOC lowering its frequency when temperature goes high, sleep over 1 minute before each benchmark to prevent this issue.
fps performance: fps = 1000 / avgtime(ms)
cpu usage: take the CPU value in top utility output
memory usage: take the RES value in top utility output
the overall power consumption and performance per watt:
Disable usb charging: adb shell echo 0 > /sys/class/power_supply/battery/charging_enabled
current(μA) = adb shell cat /sys/class/power_supply/battery/current_now (multiply -1 for 810 chip)
voltage(μV) = adb shell cat /sys/class/power_supply/battery/voltage_now
power consumption(mW) = current / 1000 * voltage / 1000 / 1000
performance per watt(1000fps/W) = fps / power consumption * 1000
The binary size after debug stripping
squeezenet
mobilnet
vgg16
