C-based/Cached/Core Computer Vision Library, A Modern Computer Vision Library

Acf feature based method of piotr dollar's toolbox get excellent performance for face detection and pedestrian detection. I appreciate you can implement this in your ccv (replace icf feature).

I'm new to ccv, so I thought I'd start by building latest stable (and unstable), and at least see if I could duplicate the bbfdetect face detection accuracy score from https://github.com/liuliu/ccv/blob/stable/doc/bbf.md

Unfortunately, I don't seem to be able to do so; I only get a score of "53.82% (30)" when I follow the steps; far worse than the quoted "82.97% (12)". I get the same result for stable and unstable branches.

I'm running on OSX 10.8.3 with the following lib/config.mk:

XCODE_SDK := macosx10.8
ARCH := x86_64
CC := clang
AR := ar
CFLAGS := -I/usr/local/include -arch x86_64 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.8.sdk -msse2 -D HAVE_SSE2 -D HAVE_LIBJPEG -D HAVE_LIBPNG -I/usr/X11/include -D HAVE_FFTW3 -D HAVE_LIBLINEAR -D HAVE_ACCELERATE_FRAMEWORK
LDFLAGS := -L/usr/local/lib -lm -arch x86_64 -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.8.sdk -ljpeg -lpng -lz -L/usr/X11/lib -lfftw3f -lfftw3 -llinear -framework Accelerate

I have local installs in /usr/local of libpng(1.5.14), libjpeg(v9), fftw(3.3.3), and linear(1.93). I also tried gsl(1.15), with same results. I have XCode(4.6.1) with up-to-date command line tools etc.

I converted the CMU gif files to png as per your example, using ImageMagick (6.8.3-10, darwin 12.3.0).

My generated CMU filelist.txt has 130 entries, and truth.txt has 511 entries. After running bbfdetect on filelist.txt, the generated result.txt has 305 entries.

Do you have any thoughts or suggestions as to how I might be able debug this further, so that I can match your results from the bbf.md doc?

Thanks.

It is great that ccv will add icf based algorithm for pedestrian detection. Hope to provide a detection model and brief usages (just like BBF, DPM, etc) to use it.

Hey I am training cifar-10.c using the cifar dataset. I want to inform the maker that I think there is an error in the backpropagate function in the ccv_convnet.c file. Plz someone help. Its urgent

While profiling some of the different built in functions I noticed that ccv_sobel() appears to consume a significant amount of time particularly on iOS devices. I would assume the best way to attempt to improve the performance would be to use SIMD instructions, like NEON for ARM or MMX for Intel. However given the way the current function is implemented it seems like it could become messy and a non-trivial task. Is this as good as it gets or is there a way to get performance up cleanly?

I am using OSX lion( 10.7.4) on a i7 quad core. When I go to the lib directory and ./configure and make (choosing clang) I am able to successfully compile from that make file. However when I move into the bin directory to compile the c files in that directory this is what follows:

➜ bin git:(unstable) make make -C ../lib

USE: clang COMPILE FLAGS: -D HAVE_LIBJPEG LINK FLAGS: -ljpeg

make[2]: Nothing to be done for `all'. clang bbffmt.c -o bbffmt.o -c -O3 -Wall -I../lib -D HAVE_LIBJPEG clang -o bbffmt bbffmt.o -L../lib -lccv -ljpeg -lm Undefined symbols for architecture x86_64: "_isnanf", referenced from: _ccv_any_nan in libccv.a(ccv_util.o) ld: symbol(s) not found for architecture x86_64 clang: error: linker command failed with exit code 1 (use -v to see invocation) make: *** [bbffmt] Error 1 ➜ bin git:(unstable)

I feel like this is just a simple flag issue in the makefile. how do i get it to recognize 64bit ness or will i have to pass a i386 flag?

Good morning, Liu Liu. I am a student and researcher of University of Campinas (UNICAMP) in campinas, Brazil.

In my latest project we worked with you ccv lib. The main purpose of the project was accelerating the execution time. In the study our main 'weapon' was the usage of parallel routines provided by open mp.

With that said, we found 2 hotspots for optimizing your code, they are: 1 - Parallelize the execution of ccv_swt (called in ccv_detect_words) for each channel (dark_to_bright and bright_to_dark). This reasulted in a good speedup! 2 - The second optimization was based on canny filter. We noticed that ccv_canny was called inside ccv_swt and that ccv_canny called ccv_sobel inside it. But ccv_sobel was also calculated in ccv_swt! So we moved the calculation of ccv_sobel in ccv_swt to before ccv_canny and created a new function based on ccv_canny called ccv_canny_no_sobel which is the same as ccv_canny but receives as input the result of a sobbel execution and don't calculate it again inside it! This also generated a good speedup!

This 2 optimizations can be found on a fork that I did of your branch. It can be found on https://github.com/krautz/ccv

If you like our research project, please feel free to let me send you a merge request with this optimizations!

Thank you! Caio Krauthamer

Hi,

I've discovered some serious problems in SWT implementation. Looks like it corrupts memory.
Here I've some added code to reproduce the problem, full description how to run it is in Readme.

As I can see, there are problems with canny implementation and caching code. Help is required to fix it.

One of the errors (valgrind output): ==6025== Invalid write of size 4 ==6025== at 0x5C2E09: ccv_canny (in /home/xxx/src/repo/libccv_fork/ccv/test/testapp/Debug/testapp) ==6025== by 0x40835E: ccv_swt (in /home/xxx/src/repo/libccv_fork/ccv/test/testapp/Debug/testapp) ==6025== by 0x4399CF: ccv_swt_detect_words (in /home/xxx/src/repo/libccv_fork/ccv/test/testapp/Debug/testapp) ==6025== by 0x4032E6: detectTextOnImage(ccv_dense_matrix_t*, ccv_swt_param_t const&, bool) (main.cpp:35) ==6025== by 0x403296: main (main.cpp:29) ==6025== Address 0x69cc790 is 560 bytes inside an unallocated block of size 162,768 in arena "client"

I'm trying to run training on a mac and tried two different computers, both fail here:

- at epoch 001 / 100 => going through 1 / 2 for testsAssertion failed: (CCV_GET_CHANNEL((*a)->type) == convnet->channels), function ccv_convnet_encode, file ccv_convnet.c, line 647. Abort trap: 6

I'm wondering if you could shed some light on how to fix this.

Thanks

Hey, I compile ccv without any issues (Ubuntu 14.04 64-bit with CUDA 6.0 SDK).

However, when I try compiling the files bin/cuda to do benchmarking, I get the following errors. I checked the paths, and they look fine, and the files in lib/cuda are getting included properly, so I'm not sure why it's not able to see the static functions.

Any insights? Or can you successfully compile the binaries in bin/cuda without any issues?

make[1]: Leaving directory `/home/fatbox/code/convnet-benchmarks/ccv/ccv/lib'
clang cwc-bench.c -o cwc-bench.o -c -O3 -Wall -I"../../lib" -msse2 -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_CBLAS -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_CUDA
nvcc cwc-bench-runtime.cu -o cwc-bench-runtime.o -c -O3 -I"../../lib" -lineinfo --use_fast_math -arch=sm_30 -D HAVE_LIBPNG -D HAVE_LIBJPEG -D HAVE_FFTW3 -D HAVE_CBLAS -D HAVE_AVCODEC -D HAVE_AVFORMAT -D HAVE_SWSCALE -D HAVE_SSE2 -D HAVE_CUDA
cwc-bench-runtime.cu(18): error: identifier "_cwc_convnet_batch_formation" is undefined

cwc-bench-runtime.cu(111): error: identifier "_cwc_convnet_convolutional_backward_propagate" is undefined

cwc-bench-runtime.cu(119): error: identifier "_cwc_convnet_full_connect_backward_propagate" is undefined

cwc-bench-runtime.cu(123): error: identifier "_cwc_convnet_rnorm_backward_propagate" is undefined

cwc-bench-runtime.cu(127): error: identifier "_cwc_convnet_max_pool_backward_propagate" is undefined

cwc-bench-runtime.cu(131): error: identifier "_cwc_convnet_average_pool_backward_propagate" is undefined

6 errors detected in the compilation of "/tmp/tmpxft_00002dd2_00000000-6_cwc-bench-runtime.cpp1.ii".
make: *** [cwc-bench-runtime.o] Error 2

I am currently trying to use dpmcreate to train a single component, no parts, HOG detector baseline, using INRIA training data.

Training seems to go wrong since it reports

 - collecting negative examples -- (100%)             
 - tune model 1 constant for -inf
 - collecting negative examples -- (100%) 
 - tune model 1 constant for nan

Which I suspect to be very wrong.

Full outputs as follows

 ./dpmcreate --positive-list ~/datasets/inria_positives_annotations_fullpath_x_y_width_height.txt --background-list ~/datasets/inria_negatives.txt --negative-count 5000 --model-component 1 --model-part 0 --working-dir ./dpm_cache --symmetric 0
with 1237 positive examples and 5000 negative examples
negative examples are are going to be collected from 1218 background images
use symmetric property? no
use color? yes
negative examples cache size : 2000
1 components and 0 parts
expected 20 root relabels, 10 relabels, 50 data minings and 1000 iterations
include overlap : 0.700000
alpha : 0.010000
alpha decreasing ratio : 0.995000
C : 0.002000
balance ratio : 1.500000
------------------------
global mean: 0.339597, & variance: 0.000022
interclass mean(variance): 0.339597(0.000022)
computing root mixture model dimensions: 5x15
root mixture model initialization corrupted, reboot
 - generating positive examples for model 0 : 1237 / 1237
 - generating negative examples for all models : 5000 / 5000
initializing root mixture model for model 1(1)
 - creating initial model 1 at 5x15
 - converting examples to liblinear format: 6236 / 6237     
 - generated 6236 examples with 2326 dimensions each
 - running liblinear for initial linear SVM model (L2-regularized, L1-loss)
.*
optimization finished, #iter = 11
Objective value = -2.434024
nSV = 1686
 - model->label[0]: 1, model->nr_class: 2, model->nr_feature: 2325
components == 1, skipped coordinate-descent to optimize root mixture model
initializing part filters
 - initializing part filters for model 1(1)
optimizing root filter & part filters with stochastic gradient descent
 - collecting responses from positive examples : 100%  
 - positive examples divided by components : 1192
 - collecting negative examples -- (100%)             
 - tune model 1 constant for -inf
 - collecting negative examples -- (100%) 
 - tune model 1 constant for nan

I stopped the training since -inf and nan indicate that something is very wrong.

Using model-part bigger than 0 generates a segfault.

I am using the stable release of libccv liuliu-ccv-70f8126.

Any idea of what is going on ?

Hi,

I am using ccvwrapper from https://zablo.net/blog/post/stroke-width-transform-swt-python/index.html. When i tried clang to make .so file on ccvwrapper it is missing required symbols. I tried using steps mentioned https://pypi.org/project/ccv/ , but the patching is not working. Can anyone assist me creating wrapper on ccv .

Thanks, Arun.

Very happy to see this ported to Metal. Now that I've given up on S4TF, it would be interesting to learn of a Swift framework (s4nnc) that runs deep learning on Metal. Regardless, I'll probably use MPSGraph directly in my future work instead of a high-level API.

I haven't seen the nnc directory of this repo get changes, but will that be ported to Metal soon? I also haven't seen Metal shader files yet, while there are custom CUDA shaders. Don't be afraid to use MSL for operations not supported by MPSGraph.

Hi

using ccv with TLD, I want to read pictures from /dev/video0 with an usb cam.

I will to using ffmpeg like this : ffmpeg -f v4l2 -framerate 25 -video_size 640x480 -i /dev/video0 output.avi I have try to follow your example on get starting like :

#include <ccv.h>
int main(int argc, char** argv)
        ccv_dense_matrix_t* image = 0;
       ccv_read(argv[1], &image, CCV_IO_RGB_COLOR | CCV_IO_ANY_FILE);

But seem not clear to me... May I have other possibilities to do that more easily? How can I do this? thank's for help. Laurent.

Hello, Is there a guide/tutorial to build ccv on Windows 10? I tried to make the library in the bin folder and it did not seem to work. Can anyone point to any specific resources?

Any help is appreciated. Thanks!

Per: https://docs.nvidia.com/deeplearning/sdk/cudnn-release-notes/rel_8.html

Significant performance regressions observed for cudnn 8 on 2080 Ti (resnet50 from 1400fps down to 300fps with 4x2080Ti).

When building the library, the compiler gets stuck on ccv_resample.c and eventually runs out of memory and dies. I have 8GB of RAM and 2GB of swap. I'm using gcc 7.4.0 on ubuntu 18.03. Does this have anything to do with the preprocessor and the "for_block" macros in the code?