Vincent Rabaud's profile - activity

that helps but it is not complete. One way is to change the behavior of the cv::Exception. But there are still some std:

compiling OpenCV without exception support Is there any way to compile OpenCV without exception support ? Even the 3.4 b

You should probably ask that question on http://answers.ros.org/questions/ Build OpenCV like you would for any source package in ROS using http://wiki.ros.org/kinetic/Installat... (but have the rosinstall_generator only mention opencv3.

For creating one lib in debug only, I don't think it's feasible through CMake. You can still build OpenCV in Release and Debug mode and copy the files over, I believe that should work.

Here are the flags used on Hydro: https://github.com/ros-gbp/opencv2-release/blob/debian/hydro/raring/opencv2/debian/rules#L20

SSE is probably the one you need to enable. Maybe there are some perf tests you can run to have a global view of your performance.

Yop, it seems that uniform_int from TR1 got renamed to uniform_int_distribution in C++11: C++11 uniform_int_distribution. Please fix and pull request Eric :)

Hi, I implemented the LSH in FLANN/OpenCV and I would concur with yes123: LSH will not give you a speedup for 1000 descriptors and probably for 10000 too. You will first need to change the parameters a lot but even then, there will be so many cache misses (as you have to go in a bunch of random buckets around) that LSH will be slower in that case.

Think about the extreme case of 1 element. With LSH, you need to compute several hash values, get the different bucks located somewhere in RAM, go over them to find the nearest neighbors and compute their Hamming distance. With your scale, most of those would be empty but the algorithm is still way more complex than a few clock ticks.

what kind of camera is it ? RGBD, stereo, mono, gray/color ?

if pcl::PointXYZ is just a struct with 3 float, just do a reinterpret_cast of the pointer to your data to a void* and then use one of the cv::Mat constructors using that.

The memory mapping between the pcl vector<pcl::pointxyz> is the same as storing 3 columns and your vector size rows.

Those parameters are the ones of the LSH multi-probe algorithm but its implementation is only for binary features right now. I'll add that to the docs. What is the type of the FREAK features ? (floats ? binary ?).

My guess is that the cast error you get is because the type is not binary.

read just gives you a frame, maybe it is a black frame ? At least you get something. What is your codec for that video ? What is the size of the array you get ? Your CMake looks fine

Thx for reporting the bug. Please use the bug tracker next time though http://code.opencv.org The bug is due to the fact that you are using gcc and C++11 flags. If you don't use those, things will be fine. It got fixed in 9018 and will be fixed in 2.4.3. Thx !

My 2 cents: I would guess that if those are not in the docs, they are meant for internal use, maybe for internal optimization sometimes. E.g., if CVCAPPROPCONVERTRGB is doing RGB conversion, it could be used internally to make sure the returned image is RGB.

Does that site help ? https://sites.google.com/site/ap186brian/activities/activity-4

when you use the VideoCapture object, the set member function is what you need to check (with the property CV_CAP_PROP_FPS).

For Android devices, I know it's a pain because there are so many cameras and I doubt they all support that feature.

What I would recommend is to multi-thread your code: have one thread that keeps calling ::grab(), and your other slower thread that calls ::retrieve() on the latest grabbed frame and then processes an image as slowly as it wants.

The speed here should not be an issue: create an image is not much compared to all the operations happening in a function.

The in-place is just faster to write and the way to go provided you can get rid of your original matrix. Each OpenCV function that does/does not support in-place should be checking for it so the code will throw if you are doing something in-place and you cannot. The documentation should also highlight that.

If you find something undocumented and that works/does not work, please file an issue on http://code.opencv.org/

method of 0 assumes no outliers. The RANSAC one requires you to give a distance and this one has to be chosen with respect to your noise model/dataset: try with a few pixels (e.g. 2) and keep increasing if you are dis-satisfied by the results.

The Least-Median method uses the median of your valid points and does not assume a distance but it assumes that at least 50% of your points are valid.