SWT error ?

[steeve]

I'm having something weird with SWT.

When I input this image:

I get this:

Would you know why that happens? (BTW, the Canny edge works fine).

[liuliu]

The SWT implementation is very much in progress. It seems to me that you set "direction" wrong, or need to upscale the image (up2x parameter). I will have a new SWT implementation based on MSER and a new stroke transform algorithm (probably in a month), which should be more stable.

[steeve]

Thanks for the quick answer! I did try with direction at 1 and -1 but got pretty much the same garbage. I'll try the up2x param, but I don't it used anywhere in the code for SWT?

[steeve]

By the way, to let you know, the image resized 2x via OpenCV's Lanczos yields the same results!

UPDATE: and also, thanks for the great work !

[liuliu]

Yeah, I forget up2x is an obsolete parameter. If assume all parameters are correctly set (as from swtdetect.c example), I suspect in this case, the direction derived from Sobel filter is off. Try use a bigger Sobel mask (or apply Gaussian filter before processing ccv_blur), if it still doesn't work, I will wait for my shining new MSER-based SWT :-)

[steeve]

Yeah I'm using the params from swtdetect.c, and I'm already doing a 3x3 gaussian blur for denoising before hand. Let me just say that when I input the whole picture (in the complete scene, not the cropped version I posted), it does work though. We'll be able to wait on your new SWT filter :)

[steeve]

Also, does SWT can function for vertical text? Like in the case above, detect the 2 lines as one rectangle? Because it's line based at the moment :)

[steeve]

BTW, what do you think happens in this case? Is a quick fix possible or this needs more in-depth fixing? Thanks :)

[steeve]

Also, I'm using clang on OS X Lion

[steeve]

Okay. I finally got to the bottom of this. It was actually caused by numpy and pointer alignment issues. Your SWT works fine! Sorry for the ordeal, but that was a nasty one to spot !

[liuliu]

Thanks for the follow up. I am planning a ccv_read function to handle these tricky memory copies. Doing raw memory operation on ccv dense matrix type is ok for me, but not going to work for other developers :-)

[steeve]

Just for future reference, here's how to convert a 2d numpy matrix (grayscale) to a ccv_matrix:

def np_to_ccv(mat):
    return ccv.ccv_dense_matrix(
        mat.shape[0],
        mat.shape[1],
        ccv.CCV_8U | ccv.CCV_C1,
        mat.ctypes.data_as(POINTER(c_ubyte)),
        0
    )