TDA4VM: OpenVX Viss kernel support

As an update,

The customer managed to instead of capturing the R, G, and B outputs, capture the C1-C4 outputs which supposedly gives me the direct outputs of the CFA. They have managed to recreate an RGB images from this, but the contrast is not acceptable. They cannot use these images for any detections and a simple bilinear interpolation works much better. Is there any way to improve this VISS node output? They have tried playing with the other parameters of the VISS node, but without luck.

(As reference for reproducibility, they used the parameters indicated in the code above, but simply captured the C1-C4 outputs instead of the RGB outputs)

Thanks,

Daniel

Hello Daniel,

The thread has been assigned to OpenVX/VISS experts based in India. Since today is a regional holiday in India, please expect a 1-2 day delay in response.

Thanks.

Hi,

If you are getting RED color output, then it should also work for other two color outputs. no reason for it not work as long as the correct image file is provided. Looking at the code above, VISS parameters seem correct, can you also share VISS create node code?

Regards,

Brijesh

Hi Brijesh,

Configuration based on C1-C4.

Creation of the inputs and outputs (C1-C3, I ignore C4). When capturing the RGB output, I would change the image type from VX_DF_IMAGE_U16 to VX_DF_IMAGE_U8 for all three images.

Finally, when I captured the C1-C3 output, I got the following images.

(raw output)

(after white balance, based on a known color strip)

As you can see, the problem here is that the color contrast is not correct. A correct image looks like this:

(before whitebalance)

(after whitebalance)

We can get a correct image from using the OpenCV debayering process (GR bayer format) in Python, but not with the VISS node. I have attached a raw image here for you to try on.

This is a binary U16 format in bayer GR (following OpenCV standard), we sometimes use GB format since this only switches the red and clue channel. Using GR will generate a blue-ish image, using GB will generate a green-ish image like above.

The following Python code can read the image, reshape is correctly, and the converts from bayer to RGB format.

Hi Daniel,

As per below table, as given in TIOVX node's documentation, the max value of C1 to C4 does not output RGB data, instead it is used to output CFA output. Can you please chanage the value of these muxes, espectially mux2 to mux4 to value 1 to get the RGB data output? 

Regards,

Brijesh

Hi Brijesh,

As per the first post, they already tried the RGB output and nothing comes out of it. Changing this makes it even worse.

As for the CFA, I assumed it was the debayering process and since our raw images are inputted as an RGB bayer format, I could extract them like this.

Thanks,

Daniel

Hi Daniel,

I checked regarding this internally and currently only YUV422 and YUV420 outputs from VISS is tested. RGB and CFA are not tested.

May I know if they could use YUV422 as output for their usecase?

Regards,

Nikhil

Hi Nikhil,

The customer also tried YUV in the past, but after conversion of the YUV to RGB, they got similar issues with the contrast.

Best regards,

Daniel

Hi Daniel,

Are the YUV422 output correct? Could you please confirm this?

Regards,

Nikhil

Hi Nikhil,

Based on what they communicated, yes.

Hi,

In that case, if the YUV422 output is correct, then there is no issue with the VISS output. Now we would have to focus on the way of converting YUV422 to RGB. Could you let me know how they did this?

Regards,

Nikhil