TDA4VM: VPAC LDC Node and LUT

Thank you very much Gang,

I will try that and get back to you if it wont work. 

Kind Regards,

Myrah Naeem

Hello Gang,

I have reversed the order of coordinates like you have mentioned in the link . I am attaching the the LUT header file i have tested and the result I obtained after changing the order. Please check the LUT i have attached. Thank you 

Original Image :

Resultant Image:

Kind Regards,

Myrah Naeem

LUT with positions according to TI format

LUT_TI_REV.h

Hi Myrah,

You did not reverse the order of H and V offsets.

"LUT_TI.h"

    0x95, 0x02, 0x43, 0x03, ...

"LUT_TI_REV.h"

    0x02, 0x95, 0x03, 0x43, ...

to try:

    0x43, 0x03, 0x95, 0x02, ...

Hello Gang,

I even tried this order but the result is not correct. I am attaching the LUT header file and resultant image i have obtained 

"LUT_TI_REVERSE"

0x43, 0x03, 0x95, 0x02, ..

LUT_TI_REVERSE.h

INPUT IMAGE

RESULTANT IMAGE 

Kind Regards,

Myrah Naeem

OK, this looks like the right encoding format at least.

You mentioned above:

>>> "I am not downsampling my LUT so I kept the downsampling factor as 1"

>>> "obj->mesh_params.subsample_factor = 1"

This "subsample_factor" is m = 0, 1, 2, ..., 8 for 1x, 2x, 4x, ..., 128x down sampling.
So, you should use 0 for no down sampling.
Typically, you should not get valid output if you use 1 instead of 0.

Please try 0 and let me know the result.

BTW, "m" shall be at least 3 or 4 for real applications and smaller values are only for debugging and testing purposes.

Thank you very much gang for helping me out. 

I kept LDC FACTOR = 0 

>>> "obj->mesh_params.subsample_factor = 0"

The output image is attached below and the error I received at the terminal was 

 VX_ZONE_ERROR :[tivxVpacLdcProcess:363] failed to submit Request

Resultant Image :

Thanks and Regards,

Myrah Naeem

Hi Myrah,

Interesting: I see how it could work with 1 in wrong geometry, but I don't understand what is wrong with 0.

If you keep it as 1 and down sample your current 1280x944 LUT by 2x2 to get a 640x472 LUT, and replace the top-left 640x472 corner of your current LUT with this down sampled one, you shall get correct geometry.
Can you try that so that we can verify your geometry first?

Best,
Gang

Hi Myrah,

What is the output block size you are using for this experiment? 

Regards,

Brijesh

Thank you Gang, I have not understood this part "replace the top-left 640x472 corner of your current LUT with this down sampled one" can you please elaborate?

Secondly, while creating the mesh image i changed value of table width and height. So I tested it again with original table width and table height and the result is valid .I am posting results of both. 

Amended (table_width, table_height values) with LDC FACTOR = 0 :

table_width = 1280;
table_height =944;                                                                                                                                                                                                                                         

obj->mesh_img = vxCreateImage(obj->context, table_width, table_height, VX_DF_IMAGE_U32);

Original (table_width, table_height values) with LDC FACTOR  = 0  :

table_width = (((obj->table_width / (1 << obj->ds_factor)) + 1u) + 15u) & (~15u);
table_height = ((obj->table_height / (1 << obj->ds_factor)) + 1u);
obj->mesh_img = vxCreateImage(obj->context, table_width, table_height, VX_DF_IMAGE_U32);

Kind Regards,

Myrah Naeem

Hello Brijesh,

I tried with these parameters. 

#define LDC_BLOCK_WIDTH (8)
#define LDC_BLOCK_HEIGHT (2)
#define LDC_PIXEL_PAD (1) 

I also put region parameters as NULL in tivxVpacNode  but it had no effect on the image.

Hello Myrah,

Can you put breakpoint on CSL_ldcCheckCfg API using CCS and see which error condition is failing? could try after checking out Gang's suggestion.

Rgds,

Brijesh

Hello Brijesh,

Sure, I will try gang's suggestion and then yours. Then i will get back to you with error.

Thanks and regards,

Myrah

Hi Myrah,

In your earlier code with "m=1" (and valid output image), LDC only used the top-left corner (640x472) of your 1280x944 LUT.
I guess it works because all sizes are correct.

output image: 1280x944
frame size: 1280x944
mesh LUT row width: 1280*4 bytes (mesh_img 1280x944)

If you replace the top-left 640x472 with your 2x2 down sampled LUT (640*4 bytes wide) and keep m=1, LDC shall work with m=1 LUT properly.
In that way, we can see if your geometry is correct.

In your above output, it looks some size is still wrong with m=0.
Since your mesh_img becomes 1296x944, you LUT shall be in 1296x944 (1296*4 bytes per row).

Best,
Gang

Hello Gang,

thank you very much gang. I have understood your point. I will implement it as soon as possible and get back to you with the results .

Kind Regards,

Myrah

Hello Gang,

I replaced top left 640x472 with my down sampled LUT with m= 1

I am attaching LUT I have created and the resultant image i have obtained. Please check it and let me know if i have implemented the way you have suggested.

LUT (with top-left replaced) :

LUT_TI_top_left.h

Resultant Image : 

Kind Regards,

Myrah Naeem

Hi Myrah,

Could you please share your original input image and original offsets with which you can get the correct output image?

Best,
Gang

Hello Gang,

This is my original input image in YUV 422 format.

original_image.zip

These are my offsets.

PCA_OFFSETS.txt

Thanks and Regards,

Myrah Naeem

Hi Myrah,

"PCA_OFFSETS.txt" has a wrong geometry if it is in LDC offset encoding format.

With the below 2 lines as examples, could you please explain how you get those 4 numbers?

    line 001:     316 69

    line 700:    -16 69

Line 1 is for output pixel at (H=0, V=0) and line 700 is for output pixel at (H=699, V=0).

What are your expected pixel locations in the input image for these 2 output pixel locations?

Best,
Gang

Hello, Myrah,

      I just have a doubt about your file, in  LUT_Ti.h , the first 2 element is :

Hello Gang ,

I have saved data in the text file that is attached "PDA_data.txt"

column 1 : x and y are initial positions of pixel,

column 2 : fix and fiy are output pixel positions,

column 3 : dx and dy are offsets which are calculated like this 

Hi Myrah,

Column 2 for fix and fiy may have issues.

1. column 2 has wrong geometry overall.
2. column 2 are all integers.
3. column 2 has a lot of out of input image boundary locations.

For example, Pixel at [0,0] in LDC output image is at [982, 78] in LDC input image.
Do you think that is correct?

    [x y] [fix fiy] [dx dy], [lut_table]
    [0 0],[982 78], [982 78],[0x00, 0x00, 0x00, 0x00]
    [1 0],[982 78], [981 78],[0x00, 0x00, 0x00, 0x00]

Hello Gang, 

Sorry I made a mistake and uploaded the wrong file. Please refer to the following one.

8055.PCA_data.zip

Kind Regards,

Myrah Naeem

Hi Myrah,

This picture above is what your output should look like given your LUT.
Is it correct?

In you "PCA_data.txt", you need to ensure "0 <= fix <= 1279" and "0<=fiy<=943" (you have some out of the image boundary).

Best,
Gang

yes gang, my output should be exactly like the image you have posted. I will put the limits and test that LUT in single camera application. Then i will let you know the results.

Thanks and Regards,

Myrah Naeem

Hello Gang,

I tried to set the boundaries but I am still not getting the exact image.

Can you please upload your pixel positions and LUT.h ?

And can you please let me know any parameters changes like pixel padding, block width and block height?

Kind Regards,

Myrah Naeem

Hello Gang,

Firstly, I tried to set boundaries and create LUT without using DCC tool to get block parameters. I used the previous block parameters but the resultant image I am getting is still not the one we require. Can you please check this code and these parameters and guide me about where I am making the mistake?

Secondly, you wrote "Adjacent LUT entries must be close by". Can you please explain what do you mean by this ? And what should be the adjacent values for out of boundary fix and fiy 

I tried 2 conditions one by one but neither of them worked and their results are same

// condition to set boundries

//CONDITION 1 

         if(fix >= 1279)
        {
        fix = 1279 ;
        } 
        
        if(fiy >= 943)
        {
        fiy = 943;
        }
        
        if(fix <=0)
        {
        fix = 0 ;
        }
        
        if(fiy <=0)
        {
        fiy = 0 ;
        }
        
//CONDITION 2

        if(( 0 <= fiy && fiy <= 943) && (0 <= fix && fix <= 1279))
        {
        fiy = fiy ;
        fix = fix;
        }
        else 
        {
        fiy = 0;
        fix = 0;
        }

After setting conditions i used the following code :

//make LUT

// find offset for x, y coordinate
dx = fix - x;     // distorted x - input x
dy = fiy - y;     // distorted y - input y

// fractional
weight_x = (8.0f * dx) & 0x7;
weight_y = (8.0f * dy) & 0x7;

qx = (int)dx;
qy = (int)dy;

xp = ((qx << 3) & 0xFFF8); // | (weight_x & 0x07);
yp = ((qy << 3) & 0xFFF8); // | (weight_y & 0x07);

//output data for one pixel
val_encode = ((xp << 16) & 0xFFFF0000) | (yp & 0xFFFF);

bilinear_table[y*outWidth + x] = val_encode;

//print 

 FILE *fp = fopen("LUT_TI.h","wt");
  
            int i = 0;

            // make header
            fprintf(fp, "#ifndef ISXO16_UB9BA_Q1_DCC_H_\n");
            fprintf(fp, "#define ISXO16_UB9BA_Q1_DCC_H_\n\n");
            fprintf(fp, "#define ISXO16_UB9BA_Q1_DCC_CFG_NUM_ELEM	(%du)\n\n", outWidth*outHeight*4);
            fprintf(fp, "#define ISXO16_UB9BA_Q1DCC_CFG \\\n{\\\n    ");

            // store items
            for (i = 0; i<outWidth*outHeight-1; i++)
            {
                fprintf(fp, "0x%02x, ", bilinear_table[i] & 0xFF);
                fprintf(fp, "0x%02x, ", (bilinear_table[i] >> 8) & 0xFF);
                fprintf(fp, "0x%02x, ", (bilinear_table[i] >> 16) & 0xFF);
                fprintf(fp, "0x%02x, ", (bilinear_table[i] >> 24) & 0xFF);
                if(i%4 == 3) fprintf(fp, "\\\n    ");
            }
            // last item
            fprintf(fp, "0x%02x, ", bilinear_table[i-1] & 0xFF);
            fprintf(fp, "0x%02x, ", (bilinear_table[i-1] >> 8) & 0xFF);
            fprintf(fp, "0x%02x, ", (bilinear_table[i-1] >> 16) & 0xFF);
            fprintf(fp, "0x%02x \\\n}\n\n#endif  /* End of #ifndef ISXO16_UB9BA_Q1_DCC_H_ */\n", (bilinear_table[i-1] >> 24) & 0xFF);
            fclose(fp);

LUT header file for Condition 1:

LUT_TI.zip

LUT header file for Condition 2:

ORI_LUT.zip

Image captured from camera :

Resultant Image (results for both conditions are same) : 

Block Parameters : 

#define LDC_TABLE_WIDTH (1280)
#define LDC_TABLE_HEIGHT (944)
#define LDC_DS_FACTOR (1)
#define LDC_BLOCK_WIDTH (8)
#define LDC_BLOCK_HEIGHT (2)
#define LDC_PIXEL_PAD (1)

Kind Regards,

Myrah Naeem

Hello, Myrah,

          I am an engineer who is also developing a similar task with you. So we can discuss in this  post.

           I read your 8055.PCA_data.zip and find 1 problem and 2 doubt 

problem 1:

for your c code to convert lut,  this line have problem 

xp = ((qx << 3) & 0xFFF8); // | (weight_x & 0x07);
yp = ((qy << 3) & 0xFFF8); // | (weight_y & 0x07);

it should be 

xp = ((qx << 3) & 0xFFF8) | weight_x ;
yp = ((qy << 3) & 0xFFF8)| weight_y ; 

otherwise, the fraction  in  your lut, is always 0  (please review your PCA_data.txt)

doubt 1;

in your  PCA_data.txt, for every point such as [0xc8, 0x02, 0xd0, 0x02],  your put offset X then offset Y.

but as Gang He's feed back 30 days ago, he mentioned  it should be Y offset then X offset.

I'm not sure which one is correct, because in the tiovx guideline, it only say. 

• 32bit encoding format (X,Y) is as follows:

so maybe you can have a try for different way 

doubt2：

your input image size is 1280*944, and LDC_DS_FACTOR = 1 , so as the tiovx guideline(TIVX_VPAC_LDC_CALC_MESH_LINE_OFFSET.), the mesh image width and height shall be 

Hello, Gang,

       I am a developer who is also meet similar problem with Myrah.

       I review this post's history and read the tiovx guideline/ sdk demo.

       I have a big doubt:

as Myrah's example,  the input image is 1280*944 and the output image size is also 1280*944.

as the tiovx guideline "he pitch/line offset for the mesh can be calculated using TIVX_VPAC_LDC_CALC_MESH_LINE_OFFSET. " and the demo code in "app_ldc_module.c",  we can see the width and height if the mesh table shall be 

Hello Gang,

Firstly, I have analyzed the whole "PCA_data.txt" and "LUT_TI.h" but I have not found the discontinuity problem or Maybe I have not understood what do you mean by discontinuity. So, can you please explain which 52 columns are discontinuous and how they should be corrected? 

Secondly, I have already sorted out boundary condition problem according to your suggestion and posted the code previously.( "0 <= fix <= 1279" and "0<=fiy<=943" )

Thirdly, To save time can you please share your LUT and data file having pixel positions and offsets?

Kind Regards,

Myrah Naeem