OAK-D S2 PoE FF blurry RGB

jamesp124

Hi,

We are considering the OAK-D S2 PoE on a UGV. I am interested in using its RGB camera for recognition of pedestrian crossing signals, which is currently done with a hand-focused 5MP machine vision camera with acceptable results.

However the OAK-D S2 with fixed-focus seems of insufficient quality. Attached are two photos taken of a pedestrian crossing. The opposite side is less than 30 meters away.

These are 4K stills requested at 1Hz over a gigabit PoE injector, so it is unlikely a bandwidth issue.

Below is the portion of the code that sets up the pipeline. Setting Lossless causes something within DepthAI to crash, so it is turned off.

cam = pipeline.createColorCamera()
cam.setIspScale(1, 1)
cam.setResolution(depthai.ColorCameraProperties.SensorResolution.THE_4_K)
ve = pipeline.create(depthai.node.VideoEncoder)
ve.setDefaultProfilePreset(1, depthai.VideoEncoderProperties.Profile.MJPEG)
# ve.setLossless(True)

cam.still.link(ve.input)
xout = pipeline.createXLinkOut()
xout.setStreamName("rgb")
ve.bitstream.link(xout.input)

cam_control_stream_name = f"{cam.getName()}_control"
control_in = pipeline.createXLinkIn()
control_in.setStreamName(cam_control_stream_name)
control_in.out.link(cam.inputControl)

Is there anything I can do to get a better image, or is this as good as it gets? (And before anyone suggests, no, bumping up sharpness to 4 doesn't magically fix it 😛)

Thanks in advance.

OskarSonc

Hey jamesp124 - to me this looks like a focus problem, since the objects closer are much sharper(if you look the part of the car on the far right)... The OAK-D S2 ideal range is 70cm - 12m so I think this is because objects are too far and the camera is FF so the objects in the far distance seem unfocused.

jamesp124

Thanks for the quick reply, Oskar! I had tested one about 9 months ago and I swear it looked much better. Maybe we just got lucky with that one.

OskarSonc

jamesp124 - was the distance maybe less than this? Or maybe you had the autofocus mode?