I run ikd-tree with a running sliding window of point cloud. pointclouds are queued in time_stamp sequential. each point are added to ikdtree and then delete later. But as I was checking kdtree points with these code: if(0) //If you need to see map point, change to "if(1)" { PointVector ().swap(ikdtree.PCL_Storage); ikdtree.flatten(ikdtree.Root_Node, ikdtree.PCL_Storage, NOT_RECORD); featsFromMap->clear(); featsFromMap->points = ikdtree.PCL_Storage; } then publish these point clouds. I found there are always points not deleted, i.e. they remain in the kdtree even though deleted.

Another question, what if i add_Points with downsample_flag=True. Then delete original point later. Will ikdtree delete this cuboid? what behavior will ikdtree generate?

Hi, thanks for your great works. ikdtree works well with one thread but not with multi thread. the output are " *** Error in `./ikdtreeLoclization': double free or corruption (fasttop): 0x00007fcd1c000b50 *** " and the test demo are:

    std::vector<PointType ,Eigen::aligned_allocator<PointType>> pointNear;
    std::vector<float> nearDis(nearNum);
    
    omp_set_num_threads(3);
    #pragma omp parallel for 
    for (int j = 0; j < lidarCurW->size() ; j++)
    {
            PointType pointJ=lidarCurW->points[j];
            ikdtree.Nearest_Search(pointJ,nearNum,pointNear,nearDis,5.0);
 }

hope for your reply, thanks

Dear @Ecstasy-EC:

Hello there Yixi, I'd like to do PR to your ikd-tree repository. Your fantastic method deserves to be nicer!

[Features]

• Separate includes files and example files
• Add two examples (Also, I follow your code format...In fact I hate your combination format of the Snake case and Pascal case lol)
  • In particular, the public function Search_Points_by_box is added. Please refer to codes.
• Update README
  • In particular, build commands are added
  • Visualized results are added

As you're an HKUSTian, so I utilized livox-loam to get the Red Bird Sundial pcd file :smile:

In addition, my large_scale_map.pcd is used for an example 3, i.e., the async example. I think this file can be substituted by any file. If you'd like to use this file, it would also be better to upload large_scale_map.pcd on your lab's official drive (e.g. Google drive). If then, please README.md appropriately!

Cheers~

Dear Admin:

First, thank you for sharing your invaluable codes.

But, I'd like to use your incremental Tree algorithm to achieve radius search; unfortunately, I think the code doesn't support the radius search function.

Is there any particular issue not implementing radius search?

Thanks in advance!

transformed KDClass into templated class for easy integration with PCL and custom point types. All external struct and classes have been moved inside KDTree class for uniformization

I've also changed the demo.cpp to conform with new usage. My next intention is to write a PCL wrapper to use as a PCL::Search object

Hi, Nice work! I found you didn't initialize the variable of "Rebuild_Ptr" in your code. When I run your demo, everything seems ok, but when I use your code into my own project with ROS it always leads to problem crash, I solved this problem after I set the initial value of "Rebuild_Ptr" to nullptr. I can't sure is this a bug or I just somehow make my program right. So Can you help me to see if this problem will cause the program to crash? Best wish.

hi Thanks for sharing your work, it is a great improvement that make kdtree cost less time when add new points, however, if I want to delete some points, can this framework handle without build a new kdtree?

你好，非常感谢分享工作！ 我有一些疑问，还请解惑： 1.一般在里程计的scan2map过程，是需要找到当前帧附近邻域内(比如50米)的历史的关键帧，也就是目前代码中的localmap(我认为)，如果里程计一直是单向在跑，我认为ikd-tree一直增量的更新这个树，确实没问题，但是一旦调头了，那ikd-tree不是已经把历史的一些点给删除了吗，所以这里会不会存在问题呢？换句话说就是ikd-tree能否保留历史关键帧上的点呢？ 2.在测试的时候我发现ikd-tree的size()一直在增大，是不是我忽略了什么细节？ 3.对这个ikdtree非常感兴趣，可否加个微信什么的方便私底下询问，我的是16621710962。万分感谢了

I did some tests and read the code. It seems that ikd-tree does not collect those points deleted in the downsample process. Can I get the missing points from "Downsample_Storage" ?

(I really need those points, because I need to do object tracking with them.)

PCL kdtree provides index to mark origin map points and all points are managered out of kdtree, but ikd-tree does not have index.

Now I define an embedded index uint32_t value in "PointType" to solve this problem, but it's not perfect because it's growing and will overflow soon.

Thanks for your work, it's elegant and useful.

I would like to use nearest search in a openmp section, eg:

` #pragma omp parallel for num_threads(4) for (int i = 0; i < laserCloudSurfLastDSNum; i++) { PointType pointOri, pointSel, coeff; std::vector pointSearchInd; std::vector pointSearchSqDis;

pointOri = laserCloudSurfLastDS->points[i];
pointAssociateToMap(&pointOri, &pointSel); 
kdtreeSurfFromMap->nearestKSearch(pointSel, 5, pointSearchInd, pointSearchSqDis); //replaced by ikd-tree

} ` Is it possible to do multiple search at the same time ? It seems that a mutex is used within the Nearest_Search(), which means that parallel operations actually save no time. Am i right ?

Hi, from what I can see in the code, you are using multiple threads, but all of the calculations are done by the CPU correct? Would it be possible to adjust the code to make use of the GPU using something like CUDA to be able to achieve even better performance?

I can't understand the line 896, Isn't it supposed to be like this “ if(q.size()< k_nearest && (dist_left_node < q.top().dist || dist_right_node < q.top().dist)” ?

I am doing some rewrite of this ikdtree to adapt my situation. And I find you used nth_element while building a tree. According to my experiments there might be a bug.

Suppose at some point division-axis is 0, which is x-axis, let's use 1-D data to explain:

Suppose we have a vector [0,1,2,2,2,2,2,2,3,4] which is points' x-axis value after nth_element() function call, y-axis value is totally different, say [0,1,2,3,4,5,6,7,8,9]。nth_element（） divide these points into different group but not determinately which point is on left or right node. point (2,2) might be on the left or right.

Thus while searching, or delete or add by point. it not guaranteed to find this point.

https://github.com/hku-mars/ikd-Tree/issues/22

Right now the header file is "using namespace std". This is considered bad practice, but easy to fix.

Additionally <unistd.h> is used to put the current thread to sleep. This is not platform independent, but easily replaced by C++ std library functionality.

Right now the header file is "using namespace std". This is considered bad practice, but easy to fix.

Additionally <unistd.h> is used to put the current thread to sleep. This is not platform independent and can easily be replaced by C++ std library functionality.