search

csp256 / cudaLATCH

GPU implementation of LATCH descriptor & matcher.
40 stars 12 forks source link

latch method call in min.cpp returns empty Descriptors. #7

Open buzzsuresh opened 8 years ago

buzzsuresh commented 8 years ago

Hi Chris,

Trying to run your Cuda LATCH implementation code.

Had fixed the broken compilation issues with min.cpp. On execution I couldn't get the descriptors after the "latch" method call. Had copied the descriptors into host and tried to print them, but it fails since there is no descriptor into d_D1.

For this reason I believe matching also fails (processing 'bitMatcher' method) which always gives me a count of 1. 

FAST(img1g, keypoints1, fastThreshold);
latch( img1g, d_I, pitch, h_K1, d_D1, &numKP1, maxKP, d_K1, &keypoints1, d_mask, latchFinishedEvent );
cudaCalloc((void **) &h_D1, sizeD);
cudaMemcpyAsync(h_D1, d_D1, sizeD, cudaMemcpyDeviceToHost);
cout << "\nDescriptors 1: " << endl;
if ( h_D1 )
    for ( int i = 0 ; i < sizeD ; i++ )
        cout << h_D1[i] <<  " ";

Looking forward for your suggestion and Please help me on this.

thanks.

csp256 commented 8 years ago

As you figured out min.cpp is broken. I believe the data format changed since I made min.cpp so your best bet is to look at vo.cpp

The only reason min.cpp existed was to remove the pipelining I added for efficient visual odometry in vo.cpp (that is, how the CPU is processing one frame while the GPU is working on another, making the GPU code zero overhead). This was only done so less experienced developers could see how to use CLATCH with a lower barrier to entry.

If you are dead set on fixing min.cpp, try taking vo.cpp and ripping things out until you have a minimally functional example. That is what I did the first time.