RetinaFace detector with C++
I convert mobilenet-0.25 mxnet model (trained by yangfly) to caffe model
I convert R50 mxnet model to caffe model BaiDuYun密码:6evh | Google Drive
-
I have checked the output of the two models be the same.
-
For same input images, the output of the two detector (python version and cpp version) is same.
(the code is too simple, only for reference 23333)
This is anchor parsing demo, NO INFERENCE, you should replace your own inference header and code (caffe/ncnn/feather .etc) in the source code
I create QQ group: 760688309 for communicating
I update time test tools for ncnn and rokidnn, only arm64-v8a version (I'm too lazy to compile armeabi-v7a version, if you need, I can update later)
how to run
use adb push to device and run ./retina_det imagepath modelpath thread_num loop_num
It will calculate average time (including input time, forward time and anchor parsing time, nms time) the output image is saved as "res.jpg"
VGA (640x480)
RokidNN vs NCNN
Next week I will provide apk built with NCNN and RokidNN for you to compare.
| platform | net | threads 1/2/4(ms) | inference |
|---|---|---|---|
| qcom625 | mnet | 418/262/202 | NCNN |
| qcom625 | mnet | 379/244/180 | NCNN(optmize) |
| qcom835 | mnet | 137.37/82.97/62.79 by hanson-young | NCNN |
| qcom835 | mnet | 125.10/74.52/75.09 by hanson-young | NCNN(optmize) |
R50: Because of the GPU memory limited, I set max(width, height) to 1000 and test it on WiderFace_val set for SINGLE SCALE, NO MULTI-SCALE, NO FLIP, results as follows:
| wider val | easy | medium | hard |
|---|---|---|---|
| python version | 92.25 | 88.86 | 64.02 |
| cpp version | 91.36 | 87.15 | 62.22 |
mnet: I test on WiderFace_val set for SINGLE SCALE, NO MULTI-SCALE, NO FLIP, results as follows:
| wider val | easy | medium | hard |
|---|---|---|---|
| python version | 83.28 | 77.02 | 39.52 |
| cpp version | 83.04 | 76.84 | 39.43 |
You can see the two detect results as results/cpp_detect and results/python_detect. The main cause of the difference in test results is the numerical precision difference between C++ and python.
- implement bbox_vote
- multi-scale test
- speed benchmark (CPU/ARM)
- new models