search

CTU-IIG / kcf

Kernelized Correlation Filter tracker
13 stars 6 forks source link

Unhelpful error message #3

Closed wentasah closed 6 years ago

wentasah commented 6 years ago

I'm trying cufft version on TX2. When I run ./kcf_vot -v ~/dataset/bag/ it gives me the following error:

Image dimension after forward FFT are too big for CUDA kernels.

If that means that I should use the -f option, please add it to the error message.

Shanigen commented 6 years ago

I added much more verbose error message, which should help more.

wentasah commented 6 years ago

What does it mean "its size is less 1024 pixels". Does it mean that width×height is less than 1024?

Shanigen commented 6 years ago

Yes, that is exactly what it means. Should I rephrase it?

wentasah commented 6 years ago

That's strange. When I just add -f, which should be the same as -f128x128 (i.e. 16384 pixels), the error message disappears. So something is wrong here.

Shanigen commented 6 years ago

But 128x128 is the size of the window before the FFT and the HOG, after the HOG and the FFT the window in the frequency domain is only 32x17, which is 544 pixels. The error message tells you that. The CUDA kernels are in the frequency domain.

wentasah commented 6 years ago

The original message told me "after forward FFT" and FFT does not change the size. But even if it does, the user might not be familiar with all internals of the software. The error message should tell the user how to make it work, not explain some internals of the software.

If I'm not mistaken, each HOG cell is 16 pixels and FFT cuts the size to half. So can you put there something like "Please use -f option with WH < 32769" (i.e. 1024162+1 = limit HOGsize * FFTreduction + 1 to use < instead of <=)?

Shanigen commented 6 years ago

I have changed the error message so its tells the user how to correct the error and also made other error message for dimensions that are not divisible by HOG cell size also more clear.