Got wrong Result

[doublexxking]

Hi. I used your code to implement my own data. I use PCL to read two point cloud from PCD files, one is Scene_nn scene; KDTree_cpu kdtree_cpu; scene.init_Scene_nn_cpu();

I also find normals from PCL and init KDTree in cup and push to GPU as init_Scene_nn_cuda does.

The other one is used by thrust::copy to get a device vector.

However, when I use cuda_icp::ICP_P2P_cuda to get results, I find the first iteration works fine, but it gives a crazy extrinsic (variable in ICP_P2P_cuda funtion). Thus, the second iteration cannot find a reasonable results and Ab[28] is 0 to return the function. The final result is the crazy extrinsic in first iteration.

I am not sure why it give me this results and I use your data works very good.

It is appreciate that you could give me some hints.

Thanks in advance.

[meiqua]

Are the point clouds roughly aligned initially? You can check this by visualizing them in open3d.

[doublexxking]

The point clouds is aligned. The PCL can give me correct results. And it is why the first iteration works fine but the second iteration always wrong.

[meiqua]

How does open3d's icp work like here?

[doublexxking]

I do not compile open3d but I change all to PCL to visualize the point cloud and check my results. After PCL, the two point clouds aligned very well..

[meiqua]

what's the fitness after the first iteration?

[doublexxking]

*first*** count: 1076.000000 fitness: 1.000000 inlier rmse : 4.325136

*second*** count: 698.000000 fitness: 0.648699 inlier rmse : 78.470566

*third*** count: 0.000000

My model is in mm, so I set max_dist_diff = 100.0f;

[doublexxking]

I also attached two images, first one is before the ICP and the scond one is after the ICP

[meiqua]

Can you paste some relevant codes? The bug is in the initial function of clouds and normals maybe.

[doublexxking]

After your hints, I find I actually got wrong normals. Thanks!!!

[aditya2592]

Hi @double344931987 . Do you have any sample code to run the ICP between a depth image form cuda_renderer in the repo and a random scene or between two random point clouds?

[aditya2592]

@meiqua In the get_normal() function in common.cpp, what units are following in and what do they signify :

 int distance_threshold = 2000;
 int difference_threshold = 5

[meiqua]

Units: mm. distance_threshold: if depth is greater than this, normals are not calculated. difference_threshold: in the sliding window, if depth - center depth > threshold, depth at this pixel will not be considered for normal.