Hello.
I'm using the pycaffe API to try and run MSCNN on a single image using the following net and pretrained model
mscnn-8s-768-trainval-pretrained/mscnn_deploy.prototxt
mscnn-8s-768-trainval-pretrained/mscnn_kitti_trainval_2nd_iter_35000.caffemodel

No matter what i do (use caffe.Net+net.forward or net.Classifier+net.predict) i end up with my memory (8Go) and swap (30Go) being saturated and the script crashing.
Manually killing the process shows it's stuck in self._forward.

In GPU mode i get the infamous Check failed: error == cudaSuccess (2 vs. 0) out of memory line before anything starts.

Did anyone get these memory issues?

EDIT : ok, seems like i clearly lack memory.

Hi @zhaoweicai,

I migrated your code to window platform(based on the happynear/caffe-windows), and ran your code as you describe in the README file. I trained the network of mscnn-7s-384-2x and mscnn-7s-576-2x.
While the performance of mscnn-7s-384-2x in you paper is 90.55 87.93 71.90(Easy, Moderate, Hard), I get 86.52 87.75 75.96(Easy, Moderate, Hard).
While the performance of mscnn-7s-576-2x in you paper is 94.08 89.12 75.54(Easy, Moderate, Hard), I get 86.96 88.33 77.96(Easy, Moderate, Hard).

I tried three time for mscnn-7s-384-2x, the results were close. The easy part was worse obviously and hard part was better. Are the results all right?

Hi, @zhaoweicai

When I tried to install the Caffe version provided in the repository, I can "make all". But when I tried to "make test", the "test_detection_layer.cpp" encountered errors as following:

CXX src/caffe/test/test_detection_layer.cpp
src/caffe/test/test_detection_layer.cpp: In member function ‘virtual void caffe::DetectionLayerTest_TestGradient_Test<gtest_TypeParam_>::TestBody()’:
src/caffe/test/test_detection_layer.cpp:85:20: error: ‘class caffe::DetectionParameter’ has no member named ‘set_field_r’
detection_param->set_field_r(2);

I checked with the original Caffe version, I noticed that this "test_detection_layer.cpp" is new. I traced back the "DetectionParameter" in "caffe.pb.cc" and there is NO "set_field_r".

Similar errors occurred when running "test_iou_loss_layer.cpp". It says that the "IOULossParameter" is undefined.

Do you have any suggestions on how to solve these problems?

Thanks.

看作者的名字感觉作者是个**人，水平太高不好意思打扰了，希望看到的国人同胞帮我看看：
论文提到测试的图像7518张是没有标签的，采用分离验证的方法，那么论文用作者训练好的模型在测试集来测试生成的标注怎么评价呢 ？
看了KITTI数据集上的评估函数devement kit 的方法，主函数发邮件是怎么回事？测试集标签可以生成吗 ？

@zhaoweicai
I've download the pretrained model of '8x_768' for test. But it loged a OOM message. I'm using 1 K40 for test, the memory is equal to Titan used in the papar.
I try to train the model by myself then, but the OOM proplem occurs again when I using the default batch size, and when I reduce the batch size by half, It seems normal but uses about 10G memory.
It's so strange..

HI @zhaoweicai
According to the caffe code you provided , I wonder that if the crop_width and crop_height in image_gt_data_param determined the input size of training net, if so , the input size of training net should be different from the net in deploy file because there is no crop in deploy, so how it works correctly?
Thank you!

hi @zhaoweicai,

I tested "ms-7s-720-pretrained" in Caltech-usa-test in reasonable and got great performance 10% (log-average miss rate) as the record in the paper.

Then I tried to train a new model based on "mscnn-7s-720". The training data is Caltech-train04 which includes 32077 images. After training, I got "mscnn_caltech_train_2nd_iter_25000.caffemodel" in the "mscnn-72-720" directory. However the performance is 0.89 (log-average miss rate). The reason may be the training data but I am not sure. Could you help me, thanks.

Would like to know if there is a python based demo instead of Matlab?

Does the 'r' denotes "range" or "rate"?
And why did you give whr and xyr a limit?
I can't find any comment on this.
Thank you very much for your help!

const Dtype min_whr = log(Dtype(1)/field_whr); //default: -1
const Dtype max_whr = log(Dtype(field_whr)); //default: 1
const Dtype min_xyr = Dtype(-1)/field_xyr; //default: -0.5
const Dtype max_xyr = Dtype(1)/field_xyr; //default: 0.5

Thanks for sharing your code!
I have read some of source code. I am confused about your split of train/val/test sets.
In the Caltech train_val.prototxt , training set is set00-05, the validation set is this

As shown in the figure, in the mscnn_window_file_caltech_test.txt , the validation images are from set06-10.
So your testing set is the same as your validation set??
I wonder whether you have upload the wrong validation sets or I have a misunderstanding about Caltech pedestrian datasets.
Hope for your reply.

Hi @zhaoweicai
I am working on a Python implementation of the testing script on the CalTech experiment. I have some questions about the network output of feedforward.

The output is a hash(dictionary) contains three items: bbox_pred,proposals_score and cls_pred
The array dimentions are N*8, N*6*1*1 and N*2 respectively.

I don't quite understand the meaning of these outputs.

bbox_pred: I think this is the bbox coordinates, but why the dimension is N*8 ? A bbox only need X1, Y1, width and height totally 4 columns for representation. What's the meaning of the other values?

proposals_score: I don't understand the meaning of this one. To my knowledge, Object Detection task only need the bbox and classification confidence.

cls_pred: The confidence of the prediction, I think one column is for background and another for a pedestrian.

Please give me some hint.
Thank you! :)

where is the code?

I wanted to update my MSCNN version in a Docker image i made month ago, so I rebuilt a new image, cloned the current MSCNN version and built it.
The thing is, a test that had no issues in the former image now fails because trainval.prototxt can't be parsed, the detection_param parameter isn't recognized.
Don't mind the unconventional filenames etc, i use more or less the kitty_car/mscnn-7s-576-2x prototxts.

Creating training net from net file: data/caffetestprotos/mscnn-kitti/trainval_1st_new.prototxt
[libprotobuf ERROR google/protobuf/text_format.cc:245] Error parsing text-format caffe.NetParameter: 382:19: Message type "caffe.LayerParameter" has no field named "detection_param".
F0130 12:30:54.677568 28 upgrade_proto.cpp:88] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: data/caffetestprotos/mscnn-kitti/trainval_1st_new.prototxt
*** Check failure stack trace: ***
@ 0x7f056d3abdaa (unknown)
@ 0x7f056d3abce4 (unknown)
@ 0x7f056d3ab6e6 (unknown)
@ 0x7f056d3ae687 (unknown)
@ 0x7f056d9f75ae caffe::ReadNetParamsFromTextFileOrDie()
@ 0x7f056db7b0db caffe::Solver<>::InitTrainNet()
@ 0x7f056db7c1ac caffe::Solver<>::Init()
@ 0x7f056db7c4da caffe::Solver<>::Solver()
@ 0x7f056db529f3 caffe::Creator_SGDSolver<>()
@ 0x40f0fe caffe::SolverRegistry<>::CreateSolver()
@ 0x408134 train()
@ 0x405b3c main
@ 0x7f056c3b7f45 (unknown)
@ 0x4063ab (unknown)
@ (nil) (unknown)

I use the nvidia/cuda:7.5-cudnn4-devel-ubuntu14.04 image as a base (so cuda 7.5 and cudnn 4).

Does someone know what changed and if it might be because of a dependancy i need to update?
Thanks.

Hi @zhaoweicai
I've successfully built MSCNN.

Currently, only Matlab version of detecting module is available, do you have any plan to write a Python version of detecting module in future? If not, I am going to write one :)

There are two training stages in proposal sub-network. The detection sub-network is based on the 1st stage of proposal sub-network, which is another stage. So, there are 3 stages in training a MSCNN?

Hi @zhaoweicai

On readme, it says that For now, only cuDNN 7.0 is supported.
But based on my knowledge, the latest version of cuDNN is cuDNN 5.1.
I guess you actually mean
only CUDA 7.0 is supported. Am I right?

As for cuDNN. what is the right version for MSCNN?

Thank you :)

The "field_h" and "filed_w" of "label_1_5x5" is 60x60.

What does the "field" mean?
It's not the anchor size in your table 1 (which is 40x40).
It's not the receptive field (which is 81x81).

looking forward to your reply!

Hi @zhaoweicai
I've successfully built MSCNN. I have got kitti_8s_768_35k_test_ped.txt. how can I evaluate by the txt document?
how can I got the average precision? (73.7%)

Hi zhaowei,

Thanks for sharing the code.I want to train models on VOC dataset, but I don't know what should I do. Thank you!

HI, I divide the training dataset into training and val datasets and use the training dataset to train my detector. How do I test the detection results on my desktop? Do I use 'evaluate_object.cpp' to test my detector performance on the val dataset? I want to find good parameters and then submit my detector's results. Thank you!

I have a question about the performance of deconvolution applied in ms-cnn.
did you check the performance of object detection with deconvolution on KITTI benchmark test images?
I want to know how much it affected your detection performance on KITTI benchmark test images. (not on validation set)
I tried to use deconvolution features in other deep learning based detection algorithms(ex) SSD: Single Shot MultiBox Detector by Wei Liu), but didn't get the expected performance improvement.

dear zhaowei,

when i run the test detection, the gpu memory is out. my gpu card has 12g memory, where i thought it's enough? i have tested the 7s and 8s modules both. could you help me figure out my problem? thank you very much

Hi @zhaoweicai

In the configuration of image_gt_data_layer, you define the min_gt_height, which is 25 for cars in KITTI when images resized to 384 * 1280. But I find that there are about 3000 cars with height smaller than 25 which are not labeled with "DontCare" or largely occluded and truncated. Why you ignore those cars?

Sorry,I forgot to set up the KITTI car dataset. I was troubled by the cudnn problem for too long

I am trying to train the MSCNN model, just as explained in readme.md, but got a segmentation fault. Is this related to GPU memory size? I am using ubuntu 14.04 with GPU grid k520 (4GB memory). How much memory is required to run the training of KITTI dataset.

....
I0131 23:47:50.833986 26703 layer_factory.hpp:77] Creating layer conv1_1
I0131 23:47:50.834022 26703 net.cpp:100] Creating Layer conv1_1
I0131 23:47:50.834034 26703 net.cpp:434] conv1_1 <- data
I0131 23:47:50.834048 26703 net.cpp:408] conv1_1 -> conv1_1
*** Aborted at 1485906471 (unix time) try "date -d @1485906471" if you are using GNU date ***
PC: @ 0x7fb0abbd9842 (unknown)
*** SIGSEGV (@0x48) received by PID 26703 (TID 0x7fb0c1a3e700) from PID 72; stack trace: ***
@ 0x7fb0d5c95cb0 (unknown)
@ 0x7fb0abbd9842 (unknown)
@ 0x7fb0ab615396 (unknown)
@ 0x7fb0ab5f1ed0 (unknown)
@ 0x7fb0abbef3ad (unknown)
@ 0x7fb0ab5f8c12 (unknown)
@ 0x7fb0ab5faf1a (unknown)
@ 0x7fb0ab5f184c (unknown)
@ 0x7fb0c3794562 (unknown)
@ 0x7fb0c03e6392 (unknown)
@ 0x7fb0c03e6f13 (unknown)
@ 0x7fb0c03e75e3 (unknown)
@ 0x7fb0c02e17c7 (unknown)
@ 0x7fb0c02e206d (unknown)
@ 0x7fb0c02e0fed (unknown)
@ 0x7fb0cde47dd9 (anonymous namespace)::opencl_fn6<>::switch_fn()
@ 0x7fb0cdc77576 cv::ocl::Context::getDefault()
@ 0x7fb0cdc776bb cv::ocl::Device::getDefault()
@ 0x7fb0cdc7771d cv::ocl::useOpenCL()
@ 0x7fb0cde506cb cv::flip()
@ 0x7fb0d73d5a15 caffe::ImageGtDataLayer<>::load_batch()
@ 0x7fb0d7352e7c caffe::BasePrefetchingDataLayer<>::InternalThreadEntry()
@ 0x7fb0d72cae85 caffe::InternalThread::entry()
@ 0x7fb0cbbefa4a (unknown)
@ 0x7fb0c66d3184 start_thread
@ 0x7fb0d5d5937d (unknown)
@ 0x0 (unknown)
Segmentation fault (core dumped)

Ok so I'm using a distributed file system and computational power and it now works with simple Caffe, using CaffeOnSpark and some adjustments. ()
But I now need to train nets for more complex tasks, including object localization.

I already used standalone MSCNN but just wanted to know if replacing caffe-public in caffeonspark by mscnn (knowing it has some differences in the data layers etc) before compiling was viable.
Has anyone done it? And if yes, do I have to tweak things before it is usable or can it be done easily?
I'm almost sure it's fine but things I didn't think about could pop up and cause some issues, you never know.

Bonus round : I'm actually using CaffeOnSpark via a Docker image so, while I don't think it's a problem, if someone knows for a fact that things can go wrong because of that, I'd be happy if you let me know!

Thanks!

Hello.
I an confused about how the mAP of the RPN is got. I tested the mAP of the RPN follow the VOC mathod but I got 49% mAP of the RPN after jointly trained.

Thanks for the nice work, as i have made some changes to the code, i'd like to check whether there are bugs in my code, thanks!

Hi, @zhaoweicai have you training models using images with different image sizes without croping?
I have met with problems like followings

detection_layer.cpp:84] Check failed: height == bottom[1]->height() (86 vs. 85)
or

image_gt_data_layer.cpp:648] Check failed: spatial_dim == label_spatial_dims[nn]

I have read your codes. I found your codes int label_height = round(template_height/(downsample_rates_[nn])); int label_width = round(template_width/(downsample_rates_[nn]));
const int label_copy_width = round(copy_width/float(downsample_rates_[nn])); const int label_copy_height = round(copy_height/float(downsample_rates_[nn]));
using simple ways to estimate the output sizes, which may not be consistent with the output sizes after conv operation.

Could you give me a help about how to train models with any image sizes? Thank you.

Hi

I've just tried to compile mscnn in a docker image. It seems that the examples folder is lacking a CMakeLists.txt file, making the build process to exit with error code 1 and crashing.

Two solutions:

• remove the addSubdirectory("examples") in the main CMakeLists.txt or
• place an empty CMakeLists.txt in the examples folder.

Best Regards

David

Hi @zhaoweicai
Lines 37~38 of testing script initalize the mu variable, then assign values [104 117 123] to it.
And in line 73, testing image subtracts mu.

I am not sure about the meaning of these operations, but I guess mu is the Mean of images.
On each testing image, we need to subtract [104 117 123] in R,G,B channels respectively.

Am I correct about this :) ?

when I trained MS-CNN , I get an error:
Check failed: fd != -1 (-1 vs. -1) File not found: mscnn_kitti_train_1st_iter_10000.caffemodel

Hi, Zhaowei

I noticed that when carrying out the training, you did not use the ground truth bounding boxes with occlusion greater and equal to 2. May I know the reason why ground truth bounding boxes with occlusion equal to 2 are excluded from the training. When occlusion equals to 2, the boxes are under the difficult category. If they are excluded from the training set, how to make sure that they can be revealed in the testing process? Did you see any drawbacks when taking those boxes into the training process?

Thanks.

Hi, @zhaoweicai ,
I would like to ask a question that what value of the ratio of negative to positive samples is used in the training of KITTI dataset. Now I set it to 6 in the training, but the trained model is very bad and there are many false positives and large location errors. Do you have any suggestions about the training?

Hi @zhaoweicai
I just found the mscnn_deploy.prototxt file for caltech in example folder is different from that in pretrained model you provided,there is no roi_cl and roi_cl_relu layer in pretrained files,is it necessary to remove these layers when I traning on caltech format data?

Hello @zhaoweicai
I was wondering if there was a reason why you didn't train cars and pedestrian together,
but trained them separately?

Thanks and have a great day~

Dear authors, is it able to merge this with the latest caffe branch so that new features could be used? Thank you.

#3702

/your/KITTI/path/training/image_2/007457.png
3
375
1242
3
1 0 551 190 580 215
2 0 605 182 638 216
4 1 424 166 506 208
5
240 178 285 214
43 184 83 214
647 171 872 203
873 172 1240 221
507 176 559 208

what is the 1 0 、2 0、４１

I am trying to retrain the model just following the guide, using the KITTI data and the label. The following error occurred and stop the training process.

I0124 22:41:18.449167 22523 layer_factory.hpp:76] Creating layer data
F0124 22:41:18.449210 22523 layer_factory.hpp:80] Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: ImageGtData (known types: AbsVal, Accuracy, ArgMax, BNLL, BatchNorm, BatchReindex, BboxAccuracy, Bias, BoxOutput, Concat, ContrastiveLoss, Convolution, Data, Deconvolution, Detection, DetectionAccuracy, Dropout, DummyData, Eltwise, Embed, EuclideanLoss, Exp, Filter, Flatten, HDF5Data, HDF5Output, HingeLoss, Im2col, InfogainLoss, InnerProduct, LRN, Log, MVN, MemoryData, MultinomialLogisticLoss, PReLU, Pooling, Power, ProposalTarget, ROIPooling, ReLU, Reduction, Reshape, SPP, Scale, Sigmoid, SigmoidCrossEntropyLoss, Silence, Slice, SmoothL1Loss, Softmax, SoftmaxWithLoss, SoftmaxWithWeightedLoss, Split, TanH, Threshold, Tile)
*** Check failure stack trace: ***
@ 0x7effabe31daa (unknown)
@ 0x7effabe31ce4 (unknown)
@ 0x7effabe316e6 (unknown)
@ 0x7effabe34687 (unknown)
@ 0x7effac494e03 caffe::LayerRegistry<>::CreateLayer()
@ 0x7effac4997ed caffe::Net<>::Init()
@ 0x7effac49b218 caffe::Net<>::Net()
@ 0x7effac5fe852 caffe::Solver<>::InitTrainNet()
@ 0x7effac5ffbe8 caffe::Solver<>::Init()
@ 0x7effac5fff39 caffe::Solver<>::Solver()
@ 0x7effac47ee53 caffe::Creator_SGDSolver<>()
@ 0x408c82 train()
@ 0x405a88 main
@ 0x7effab130f45 (unknown)
@ 0x406107 (unknown)
@ (nil) (unknown)
Aborted (core dumped)

Could anyone check what is wrong? I compiled (with cuDNN 3 as advised) and installed the code without any problem before doing the training.

```

// clipping
roi_start_w = max(roi_start_w,0); roi_start_h = max(roi_start_h,0);
int img_width = round(width_/spatial_scale_);
int img_height = round(height_/spatial_scale_);
roi_end_w = min(img_width-1,roi_end_w);
roi_end_h = min(img_height-1,roi_end_h);

here, why `width_ / spatial_scale_`?

Environment: caffe ubuntu 14.04 cuda 8.0 GTX1080 cudnn v5.
I know cudnn v5 is not recommended here but v3 is not widely used now.

Under this environment, I replaced some cudnn relevant .hpp .cu and .cpp file as the modification in the faster-rcnn. Finally, I successfully make and make matcaffe.

However when I am testing the demo, I encountered the following mistake.

I0117 19:48:49.108386 26508 net.cpp:150] Setting up label_4_5x5_data_7_split
I0117 19:48:49.108397 26508 net.cpp:157] Top shape: 4 6 9 12 (2592)
I0117 19:48:49.108407 26508 net.cpp:157] Top shape: 4 6 9 12 (2592)
I0117 19:48:49.108414 26508 net.cpp:165] Memory required for data: 26490268
I0117 19:48:49.108422 26508 layer_factory.hpp:76] Creating layer conv1_1
I0117 19:48:49.108443 26508 net.cpp:106] Creating Layer conv1_1
I0117 19:48:49.108451 26508 net.cpp:454] conv1_1 <- data
I0117 19:48:49.108464 26508 net.cpp:411] conv1_1 -> conv1_1
*** Aborted at 1484653729 (unix time) try "date -d @1484653729" if you are using GNU date ***
PC: @ 0x7f2a6b55eefe caffe::CuDNNConvolutionLayer<>::LayerSetUp()
*** SIGFPE (@0x7f2a6b55eefe) received by PID 26508 (TID 0x7f2a6be5a7c0) from PID 1800793854; stack trace: ***
@ 0x7f2a5a37c330 (unknown)
@ 0x7f2a6b55eefe caffe::CuDNNConvolutionLayer<>::LayerSetUp()
@ 0x7f2a6b58d85c caffe::Net<>::Init()
@ 0x7f2a6b58e905 caffe::Net<>::Net()
@ 0x7f2a6b4238ca caffe::Solver<>::InitTrainNet()
@ 0x7f2a6b4249dc caffe::Solver<>::Init()
@ 0x7f2a6b424ce9 caffe::Solver<>::Solver()
@ 0x7f2a6b418823 caffe::Creator_SGDSolver<>()
@ 0x40ef9e caffe::SolverRegistry<>::CreateSolver()
@ 0x4082db train()
@ 0x405f41 main
@ 0x7f2a59fc8f45 (unknown)
@ 0x4066fd (unknown)
@ 0x0 (unknown)
Floating point exception (core dumped)

I tried the original matlab version and the python version. They all crashed at the same place. I also tried the training and was also report this error.

Anybody succeed in cudnn v5? can you give me some advice?

Thanks a lot!

Hi
Can you please explain how to set up the kitti dataset and which part exactly from it should I include for the directory in the root_dir = '/your/kitti/path/ ' ??
Thank you so much for your help.
Bests,

I have tried to train kitti_ped_cyc/mscnn-7s-576-2x
I success with trainval_1st.prototxt. But when run training with trainval_2nd.prototxt. I got message:

F1115 07:35:13.703055 24222 math_functions.cu:420] Check failed: status == CURAND_STATUS_SUCCESS (201 vs. 0)  CURAND_STATUS_LAUNCH_FAILURE
*** Check failure stack trace: ***
    @     0x7f9ebe5b9daa  (unknown)
    @     0x7f9ebe5b9ce4  (unknown)
    @     0x7f9ebe5b96e6  (unknown)
    @     0x7f9ebe5bc687  (unknown)
    @     0x7f9ebedc1d18  caffe::caffe_gpu_rng_uniform()
    @     0x7f9ebed92482  caffe::DropoutLayer<>::Forward_gpu()
    @     0x7f9ebebdf661  caffe::Net<>::ForwardFromTo()
    @     0x7f9ebebdf9d7  caffe::Net<>::ForwardPrefilled()
    @     0x7f9ebed6bc89  caffe::Solver<>::Step()
    @     0x7f9ebed6c695  caffe::Solver<>::Solve()
    @           0x408393  train()
    @           0x405b61  main
    @     0x7f9ebd116f45  (unknown)
    @           0x40631d  (unknown)
    @              (nil)  (unknown)

I am using CUDA 7.5 and CUDNN 3.0.8.

Do anyone have any solution?

Thank in advance.

Hi @zhaoweicai
It's great to see MSCNN is released!

I wonder if the experimental part is going to include CalTech Pedestrian data or not? since it's also part of the experiments on paper.

I use Titian X GPU which has 12GB video memory , and Intel(R) Xeon(R) CPU E5-2650 v2 @ 2.60GHZ which has 128GB memory.
However, when I run the mscnn-8s-768-trainval-pretrained model using the run_mscnn_detection.m, it occurs "Check failed: error == cudaSuccess (2 vs. 0) out of memory" when run into the code line "outputs = net.forward({test_image}); pertime = toc;"

@zhaoweicai , could you help me how to solve this problem? thanks a lot.

hi, @zhaoweicai ,

I think one function is missed in the matlab function 'mscnn_kitti_car_window_file.m'. There is no 'readLabels(label_dir, imgidx)' function being defined.

ubuntu 14.04+cuda7.5+cudnn v5+caffe
I installed caffe and run caffe example successfully,but now I build mscnn,an error occurred,what is wrong?
error information:
In file included from ./include/caffe/util/device_alternate.hpp:40:0,
from ./include/caffe/common.hpp:19,
from ./include/caffe/blob.hpp:8,
from ./include/caffe/data_transformer.hpp:6,
from src/caffe/data_transformer.cpp:8:
./include/caffe/util/cudnn.hpp: In function ‘void caffe::cudnn::createPoolingDesc(cudnnPoolingStruct**, caffe::PoolingParameter_PoolMethod, cudnnPoolingMode_t*, int, int, int, int, int, int)’:
./include/caffe/util/cudnn.hpp:124:41: error: too few arguments to function ‘cudnnStatus_t cudnnSetPooling2dDescriptor(cudnnPoolingDescriptor_t, cudnnPoolingMode_t, cudnnNanPropagation_t, int, int, int, int, int, int)’
pad_h, pad_w, stride_h, stride_w));
^
./include/caffe/util/cudnn.hpp:12:28: note: in definition of macro ‘CUDNN_CHECK’
cudnnStatus_t status = condition;
^
In file included from ./include/caffe/util/cudnn.hpp:5:0,
from ./include/caffe/util/device_alternate.hpp:40,
from ./include/caffe/common.hpp:19,
from ./include/caffe/blob.hpp:8,
from ./include/caffe/data_transformer.hpp:6,
from src/caffe/data_transformer.cpp:8:
/usr/local/cuda/include/cudnn.h:799:27: note: declared here
cudnnStatus_t CUDNNWINAPI cudnnSetPooling2dDescriptor(
^
make: *** [.build_release/src/caffe/data_transformer.o] 错误 1

hello @zhaoweicai
On your paper you mentioned that detection speed reaches up to 10 fps in KITTI original images(1250x375). I have checked and found give or take 100ms~200ms inference time per frame depending on the model (384/576/768 + 2x + c).
If there were room for improvement what part do you think takes the most time in inferencing?

Hi @zhaoweicai

I found that bbox coordinates are represented as integers in window files.
I am wondering if it is possible to use floating number for better precisions.

Maybe this question is out of the scope, but I can't find the standards or information about the window files.
Please give me some hint :)

I find the BoxOutputLayer cost more time than other layers，and I find the BoxOutputLayer only support cpu model。Does the auther support gpu model。