webcam recommendation

[kaishijeng]

I want to try openpilot with webcam on PC
Any recommendation of webcam other than C920/C615?
Do I need to do any calibration if I use C920/C615?

Thanks,

[dorkmo]

not many people have tried the webcam option, you'll likely have to experiment and find out

[dorkmo]

feel free to leave notes on the wiki page https://github.com/commaai/openpilot/wiki/Webcam

[VirtuallyChris]

We used a Logitech C920 I believe.

[kozuch]

I guess OP may use uncompressed video stream. Try USB 2.0 or 3.0 camera modules from AliExpress (Kayeton brand etc.) They usually have YUV or YUY2 in the specs. They have m12 lens mount. Use lens that is wider than what OP needs and then convert (remap) or crop the image using opencv to correct focal length that OP expects. There are few focal length values in pixels in OP code and that is what you should end up with.

The problem with OP is that the driving model is (most likely) trained on camera images from mobile phones. Then for inference the model will work the best with image that has similar properties as the phones (lens distortion, resolution etc.) Also dynamic range, exposure etc. but you won't be able to control those on different image sensor.

But i did not really look into OP's code and the above is only my guess. Maybe there is more to this problem. The logitech cams they used are not recommend IMHO. They used them because they did not have anything better (read usb cam module with m12 lens). Also OP is said to use only about 12% of the image frame (no idea if for DNN training or for inference) - again look at the code for details. I am pretty sure you can use arbitrary camera and get very good inference results if you do it right (focal length etc.).

[kaishijeng]

Does it need to calibrate camera for a new camera matrix if I don't use C920?
There is no info of this, but I see code in warp_vis.py below:

eon_focal_length = 910.0 # pixels
eon_dcam_focal_length = 860.0 # pixels

webcam_focal_length = -908.0/1.5 # pixels

eon_intrinsics = np.array([
[eon_focal_length, 0., 1164/2.],
[ 0., eon_focal_length, 874/2.],
[ 0., 0., 1.]])

eon_dcam_intrinsics = np.array([
[eon_dcam_focal_length, 0, 1152/2.],
[ 0, eon_dcam_focal_length, 864/2.],
[ 0, 0, 1]])

webcam_intrinsics = np.array([
[webcam_focal_length, 0., 1280/2/1.5],
[ 0., webcam_focal_length, 720/2/1.5],
[ 0., 0., 1.]])

[kozuch]

Have no time for deeper analysis now but I guess intrinsics are the result of monocular camera calibration. So yes you should calibrate your new camera and plug in your intrinsics. Have a look at further code what they do with the image. They may crop it or use whatever ROI but probably if you plug in your calibration that further image processing should be done automatically i guess.

[MayDGT]

Have you succeed？

[vrucisnovi]

ne