Suggested settings for RGB and depth dataset

Hello everyone,

I just joined this forum, first of all I would like to say a big THANK YOU for this simulator.

I’m collecting images using rgb and depth sensor as stereo camera rig and I really would like to acquire images that are as much as possible close to the realistic condition.
I’m curious to know if there are suggested settings to acquire high quality images from RGB sensors. I’m already a bit familiar with CARLA and I’m already collecting the images but I noticed some artifacts under certain conditions (repetitive patterns as you know are tricky to be handled by computer vision algorithms):

Actually I just changed my settings in the Nvidia control panel to emphasizing the quality, and I know that I can apply transfer learning to enhance the dataset I generated in Carla, but I’m maybe missing some out-of-the-box solution or recommended settings.

Thank you in advance for the help.

Hi,
Although I haven’t looked into this directly - I think the artifact above is related to Unreal’s 4.24 post effects stack (I bet related to antialiasing & how motion vectors are being generated).

A few other ideas re: highest quality imagery:

  1. It is worth experimenting with supersampling images out of Carla - then downsampling images to their intended use. (although not performant, you are essentially generating more information).

  2. Force LOD_0 everywhere - again, you’re going to take a performance hit, but you’ll get higher quality images and have fewer artifacts that present themselves when objects are transitioning between LODs.

  3. Although on a different platform (unity), I’ve had good results with turning off post effects and supersampling sensor data. We used to output 4k images & downsample to 1080 which yielded the best ‘looking’ results.

I’m not certain if the above-mentioned process introduces any sort of issues related to simulator determinism. It could be argued that using an additional tool (to process/downsample) outside of Carla/Unreal adds a layer of un-trackable complexity to the mix.

Hope this is helpful.
d

Hi,

thank you so much for your reply. I had the same idea after I wrote this post and I tried to save the images in 4K but it takes too much time, unfortunately I don’t have it that much :frowning_face: I found a compromise, I switched from 1200x600 to 2400x1200. After that, I suppose I need to resize the images using for example OpenCV, and for sure I need also to consider the appropriate camera intrinsic.

Regarding the LOD I truly have no idea where I can set this and what does the 0 specifically means, have you any reference for this? Thank you in advance.

I tried to turn off the post-effects using:
cam_rgb_bp.set_attribute(‘enable_postprocess_effects’, ‘False’)

and below the different results I got on vs off (this is not the same view but the main differences are clearly visible):

Disabling it the result is much worse, but thank you for the suggestion!

I add the following image (unfortunately I cannot add two images in the same reply because I’m a new user, sorry!) I acquired today doubling up the resolution to 2400x1200. I’m trying now to maximize every graphic settings to see if I’m able to get rid of these artifacts: