AM62A7: purple edge problem on HDR3 sensor

Hi Zhili,

Sorry about the delay in replying.

For your share raw images in the previous thread, are the knee points same as below from some other your X3C thread?

0 0
1023 1023
1279 2047
1535 4095
2303 16383
2559 24575
2943 49151
3071 65535
3327 131071
3455 196607
3519 262143
3647 524287
3775 1048575
3903 2097151
3999 4194303
4095 16777215

The cause of the purple edge around the ceiling lights is also the same as in older discussions.

Around the ceiling light, the original raw image pixels smoothly rise and saturate.
The plot below is for raw image before GLBCE in log2 scale.

While GLBCE does the local tone mapping, it creates a ripple around the light boundary.

The dark pixels from the ripple cause a dark or purple band along the lights.
It is related to GLBCE handling sharp intensity changes and maintaining color.
I have not been able to find a good way to get GLBCE to suppress the ripple effectively yet.

No, the knee points used for HDR3 mode is not the same with previous one, which is shown below:

0    0
256    511
512    1023
768    1535
1024    2047
1280    4095
1536    8191
1920    16383
2304    32767
2688    65535
3072    131071
3456    262143
3712    393215
3840    524287
3968    786431
4095    1048575
65535    1048575

wang zhili said:

No, the knee points used for HDR3 mode is not the same with previous one, which is shown below:

Thanks for sharing!

Behavior is similar.

It is slightly better for the light bulb I looked at, but same in nature.
We would need to see if tuning GLBCE parameters may make any improvements.

Hi Gang,

Thank you for the analysis. I want to know how these drawings are drawn. Can you explain?

Gang Hua said:

We would need to see if tuning GLBCE parameters may make any improvements.

Do you have any suggestions for the specific parameters and direction of tuning?

wang zhili said:

Thank you for the analysis. I want to know how these drawings are drawn. Can you explain?

I just plotted the log2 values of raw RGGB pixels as a mesh in matlab around the light bulb.

The raw images is close to monochrome there.
So, it looks like a smooth surface in the plot.

wang zhili said:

Do you have any suggestions for the specific parameters and direction of tuning?

I would like to first experiment with the second pole of asymmetry LUT once I get some time.

If I plot 2 rows of log2 of raw pixels in that same area before and after GLBCE, it looks like below.
I guess we would not see any bad artifacts or color if the 2nd plot does not show any ripple.

One possibility is to saturate these areas in input raw so that no ripple is created, but that is not good for ADAS because of the traffic lights.

Gang Hua said:

I would like to first experiment with the second pole of asymmetry LUT once I get some time.

Is there any progress on this matter? Please let me know if you have any.

Hi Zhili,

I don't get any time or anyone to help in looking into this issue so far yet.

Will update you if we have any progress.