sample_camera_gmsl_raw dark & dull compared to sample_camera_gmsl (Solved)

During a direct comparison of the RAW+dwSoftISP vs the RGB ‘processed’ pipeline, we notice that image outputs are very dark and dull when using the raw sample code.

This is for identical camera type parameters given to both samples. Does the Software ISP not replicate similar processing of the RAW data to pleasing, well-contrasted RGB8 images? Our experience is dark low-contrast images with a slight pink taint.

Dear garyxiaopened,
Could you please check this symptom on our latest PDK(Drive 1.0). If you are DW 1.2 already, can you please share the camera details, commands along with screenshots of captured images.

Thanks, we’re upgrading now, were on 0.67. Camera type was ar0231-rccb-ss3322. Were there known issues with the old dwSoftISP RAW pipeline?

We’ve updated to Driveworks 1.2. The results are still very dark. The command lines were,

sample_camera_gmsl --camera-type=ar0231-rccb-ae-ss3322 --camera-port=a

https://drive.google.com/file/d/10xJofs8ZCtezeiiO1B-BC-IhXWCWYQmE/view

sample_camera_gmsl_raw --camera-type=ar0231-rccb-ae-ss3322 --camera-port=a

https://drive.google.com/file/d/11UMC7MS0HOm4zijFLpdatX8WOKWQBho8/view

The dynamic range of the software ISP processed image is very dark and does not use the dynamic range available in RGB8 images.

Dear garyxiaopeng,
Can you file a bug for with all related information?
Please login to https://developer.nvidia.com/drive with your credentials. Please check MyAccount->MyBugs->Submit a new bug to file bug.
Please share ID here to follow up. Thanks.

Tracking now on Bug #2421198. Thanks

Dear garyxiaopeng,
The sample_camera_gmsl_raw is expected to be a bit dull compared to sample_camera_gmsl. The softISP tries to mimic the HW ISP. But softISP is designed for machine vision whereas HW ISP for human.

Thanks for digging up this answer. I’d come to a similar guess and it makes sense. The difficulty is that it then seems very challenging to create images for human’s to label but for machines to learn on using this pipeline. Outdoor road scenes particularly, are regularly lacking in enough contrast to for human labellers to work with conveniently when processed via the dwSoftISP.

The ideal feature set is for RAW recording of images followed by the ability to process it according to a human or machine vision pipeline using the dwSoftISP API’s. Is any feature like this expected?

Dear garyxiaopeng,
Did you check using recording tool provided with DW for data collection?

We did not. Does it handle RAW data differently to the sample programs? We are integrating with our own data collection software built against Driveworks.

Dear garyxiaopeng,
The recording tool is to collect data and can be replayed using sample_camera_replay sample.