out of shear luck, the python inference code I wrote gave me significantly better detection results. So I did a big more debugging on the C++ netforward code under examples/yolo_9000_test.

I found one bug in yolo_v2_output.h in getResult function where the class label scores are compared. The best_score variable is currently declared as int. But it must be float. Otherwise any score less than 1 will treated as 0 and wrong class of object will be selected.

When I first tried to compile it wih cmake, there were a problem occurred like 👍

/usr/include/c++/4.8/bits/cpp_type_traits.h(314): error: expected a ">"

/usr/include/c++/4.8/cmath(80): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(80): error: expected a ";"

/usr/include/c++/4.8/cmath(105): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(105): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(105): error: expected a ";"

/usr/include/c++/4.8/cmath(124): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(124): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(124): error: expected a ";"

/usr/include/c++/4.8/cmath(143): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(143): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(143): error: expected a ";"

/usr/include/c++/4.8/cmath(162): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(162): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(162): error: expected a ";"

/usr/include/c++/4.8/cmath(183): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(183): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(183): error: expected a ";"

/usr/include/c++/4.8/cmath(202): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(202): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(202): error: expected a ";"

/usr/include/c++/4.8/cmath(221): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(221): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(221): error: expected a ";"

/usr/include/c++/4.8/cmath(240): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(240): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(240): error: expected a ";"

/usr/include/c++/4.8/cmath(259): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(259): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(259): error: expected a ";"

/usr/include/c++/4.8/cmath(278): error: inline specifier allowed on function declarations only

/usr/include/c++/4.8/cmath(278): error: variable "std::constexpr" has already been defined

/usr/include/c++/4.8/cmath(278): error: expected a ";"

After I searched I found that I need to add "-std=c++11" in cmake/Cuda.cmake line 150. With this change, I successfully compiled it.
" list(APPEND CUDA_NVCC_FLAGS -Xcompiler -fPIC -std=c++11)
"

Hope this can help someone

Is here any method to calculate the mAP of model?
And i did run netforward.cpp using 30000 iter weight, but when i checked the differences between ground truth and prediction. Does it need to be trained more?
Thanks.

Hello,
Thanks for sharing this code. I have few small questions:

1. What is the Map of your pretrained model?
2. Did you use mean image to train? if so where can i find it?
   Thanks ahead.

I have trained the model as instructions, and run the code for testing VOC image, but it has poor result.
After I learn the darknet yolo v2 source code, I find that the box_data_layer may be wrong, because it still takes side_side_(1+1+1+4), the same with yolo v1, that means one cell still only have one label class.

I have modified the box_data_layer implementation according to yolo v2 based on darknet. But it has some link problems.

When I complie caffe， I get some error, the output information is list bellow! Can you help me?

#######################################################################
In file included from /usr/include/c++/4.8/algorithm:62:0,
from /home/zouhongwei/workspce/caffe-yolo-9000/include/caffe/blob.hpp:4,
from /home/zouhongwei/workspce/caffe-yolo-9000/include/caffe/caffe.hpp:7,
from /home/zouhongwei/workspce/caffe-yolo-9000/examples/yolo_9000_test/netforward.cpp:1:
/usr/include/c++/4.8/bits/stl_algo.h: In instantiation of ‘_RandomAccessIterator std::__unguarded_partition(_RandomAccessIterator, _Ra ndomAccessIterator, const _Tp&) [with _RandomAccessIterator = yolo_v2::DATA*; _Tp = yolo_v2::DATA]’:
/usr/include/c++/4.8/bits/stl_algo.h:2283:70: required from ‘_RandomAccessIterator std::__unguarded_partition_pivot(_RandomAccessIte rator, _RandomAccessIterator) [with _RandomAccessIterator = yolo_v2::DATA*]’
/usr/include/c++/4.8/bits/stl_algo.h:2315:54: required from ‘void std::__introsort_loop(_RandomAccessIterator, _RandomAccessIterator , _Size) [with _RandomAccessIterator = yolo_v2::DATA*; _Size = long int]’
/usr/include/c++/4.8/bits/stl_algo.h:5461:36: required from ‘void std::sort(_RAIter, RAIter) [with RAIter = yolo_v2::DATA*]’
/home/zouhongwei/workspce/caffe-yolo-9000/examples/yolo_9000_test/yolo_v2_output.h:97:51: required from ‘std::vector<yolo_v2::DATA> yolo_v2::getResult(std::vector&, std::vector&, int, float, float) [with int OBJECT_NUM = 5; int SIDE = 13]’
/home/zouhongwei/workspce/caffe-yolo-9000/examples/yolo_9000_test/netforward.cpp:86:86: required from here
/usr/include/c++/4.8/bits/stl_algo.h:2245:19: error: passing ‘const yolo_v2::DATA’ as ‘this’ argument of ‘bool yolo_v2::DATA::operator <(yolo_v2::DATA)’ discards qualifiers [-fpermissive]
while (__pivot < *__last)
^
make[2]: *** [examples/CMakeFiles/netforward.dir/yolo_9000_test/netforward.cpp.o] Error 1
make[1]: *** [examples/CMakeFiles/netforward.dir/all] Error 2
make: *** [all] Error 2

######################################################################

The get_region_box function in yolo_v2_loss_layer.cpp doesn't appear to be using sigmoid to compute box.x and box.y. I wonder why? The logistic_activate() function is defined by never called.

    //why aren't we use sigmoid for box.x and box.y?
    box.x = (i + x[0 * SIZE]) / side;
    box.y = (j + x[1 * SIZE]) / side;
    box.w = exp(x[2 * SIZE]) * biases[2 * n];
    box.h = exp(x[3 * SIZE]) * biases[2 * n + 1];

I checked the ground truth label conversion steps. The ground truth values seem to be used just as normalized x, y values relative to image size. There doesn't appear to be any conversion in box_data_layer.cpp either. So how can we use the prediction value without applying sigmoid, in the mix with a linear ground truth to compute the diff?

Is it because the way the diff gradient formula is written in delta_region_box()?

.build_release/tools/convert_box_data.o：在函数‘main’中：
convert_box_data.cpp:(.text.startup+0x858)：对‘caffe::ReadBoxDataToDatum(std::string const&, std::string const&, std::map<std::string, int, std::lessstd::string, std::allocator<std::pair<std::string const, int> > > const&, int, int, bool, std::string const&, caffe::Datum*)’未定义的引用
collect2: 错误： ld 返回 1
make: *** [.build_release/tools/convert_box_data.bin] 错误 1

Could you share the yolov2 test file?
thank you very much

According to Caffe Implementation, The BN is implemented as 2 separated layers. And In your code, only the first part of BN is presented. Are you sure this is right?

Q1: Some tips occured to me when using 'make all -j8', are you annotated with the above code or the following code???
src/caffe/layers/yolo_v2_loss_layer.cpp:113:1: warning: "/" within comment [-Wcomment]
/ **/

yolo_v2_loss_layer.cpp:113
/**/ diff[index + 2 * SHIFT] = scale * (pred_box.w - truth_box.w) * pred_box.w; diff[index + 3 * SHIFT] = scale * (pred_box.h - truth_box.h) * pred_box.h; /** / diff[index + 2 * SHIFT] = scale * (log(pred_box.w / biases[2 * n + 0]) - log(truth_box.w / biases[2 * n + 0])); diff[index + 3 * SHIFT] = scale * (log(pred_box.h / biases[2 * n + 1]) - log(truth_box.h / biases[2 * n + 1])); /* **/
Q2：How to detec images through the voc_model_iter_30000.caffemodel you provide?

I have anchor boxes of my dataset, how to set anchor boxes in your code?
thanks!

I have trained the model as instruction. Then, I use NetScope to draw the network structure diagram,but it's poor error in the reorg layer, the diagram as shown below.

After I learn the darknet yolo v2 , I think that the new diagram is more more reasonable.