TDA4VM: bias of SDE disparity

Hi,

We are checking with our experts and get back as soon as possible.

Regards,

Brijesh

I want to better understand what you are observing. You plot the SDE disparity error (difference between SDE disparity and groundtruth), ant it peak at -0.8 pix while you are expecting peak at 0?

It may come from pixel locking, but I don't think it is. As far as I know, SDE disparity does not show such pixel locking problem always. 

Can you provide more information? For example, apply another disparity engine, plot the same one and compare. Even though it is less probable, there might be an issue in rectification.

SDE is HWA, so there is no way to fix it. Need to add post-processing to resolve the issue.

hi,

“You plot the SDE disparity error and it peak at -0.8 pix while you are expecting peak at 0“ ==> yes, exactly
"apply another disparity engine , plot the same one and compare" ==> yes , we use cnnstereo engine with the same calib param and imgs as SDE， and compare disparity error. The result is as follows：compared to cnn-stereo , SDE disparity error peak from -0.6pix to -1pix within (20m, 50m）.

cnn-stereo disparity result is verified as more close to groundtruth. So we want to know there is any posisbility that params or settings for SDE may impact disparity bias.

The scene : daytime, the usual road sceneirio， observed depth of ground and objects at more than 40m is obvious under-estimated。

Looking forward to your replay.

Best regards.

Sorry for late reply. 

I guess this might be artifact of the cost compression in SDE. And I don't think parameter change will improve it. I don't have clear idea yet how to resolve it, but it should be a software workaround.  DJI used to apply a proprietary post processing algorithm to SDE output disparity, and the post-processed disparity looked so good. Have you tried the post-processing to SDE output disparity?  I am curious if the post-processing can improve it.   

Thanks for your candid reply.

I want to know the normal value for bias of SDE disparity.  Will there be any improvement measures for bias?

I got your left / right images (KITTI 2012). I can create disparity map for the left image using the internal simulator. Then can you measure bias again using KITTI ground truth? It would be more accurate than transposing and reversing image.

What is the output .bin file format and your SDE config params?

Hi,

The output  data format is written by pixels, "fwrite(disparity, width*sizeof(uint16_t), height, fp);"

The configuration file app_stereo.cfg is attached

Best regards.

https://e2e.ti.com/media/dji/系统/Ti_sde_data/sde_bias/app_stereo.cfg

app_stereo.cfg

"disparity" here is of SDE disparity format, e.g. 1 sign bit + 7-bit integer disparity + 4-bit fractional disparity + 3-bit confidence? 

Hi Do-Kyoung,

Yes. Disparity data format: 1 sign bit + 7-bit integer disparity + 4-bit fractional disparity + 3-bit confidence

Hello,

I shared SDE disparity with left base image. You can download it from https://tidrive.ext.ti.com/u/wQwKHAkMChPTDAoc/bfb81572-edfe-458b-b899-3dcdf4fd392c?l

Can you measure bias again? Please let me know if binary format is not in the format you need.

Hi Do-Kyoung,

The bias measurement has been completed. Binary format is correct.

Hello,

The bias measurement result is shown in the figure, it is still bias, and it does not seem to improve.