Tail of events behind moving edge

[Moritz Beck]

Hello,

I've got a question concerning a phenomenon with the DAVIS346, that I found recently. It seems that the trailing edge of a moving object in the image has a tail of events, as shown in the figure below. In this case it was a white table tennis ball in front of a dark background moving in the right direction. Is there any explanation why only the trailing edge has this tail of events but not the leading edge?

I tried another scenario with a rotating fan and recognized the same phenomenon. Even though the rotor blades where black and the background white (so the intensity gradient is the other way around), only the trailing edge had a tail of events. All example data is filtered with the BA and hot pixel filter from the DV-Software.

I recognized similar phenomenons in papers such as "Event-Based Motion Segmentation by Motion Compensation" by Stoffregen et al. (2017), in figure 4 (fan and coin example).

It would be very helpful to understand this phenomenon. Thank you very much,

Moritz Beck

[Tobi Delbruck (INI)]

This is an interesting question that has not been fully answered yet. We
are working on it now in the context of understanding noise and dynamics
of DVS pixels.

If you want to reduce this effect, crank up the PR and SF biases and
probably also increase the event threshold. You can monitor the trail in
real time using the jAER Oscilloscope filter. See last video of this
collection:
https://www.youtube.com/playlist?list=PLVtZ8f-q0U5hD9KOM4OZ1lixhwupj9uOm
and here is the src code
https://github.com/SensorsINI/jaer/blob/master/src/net/sf/jaer/eventprocessing/filter/Oscilloscope.java

[Chenghan Li]

Hi Mortiz, thank you for the interesting observation. I can't think of an explanation yet. But if possible, could you post some photos of the exact scenes (the same object, background and lighting)? Maybe it could provide some clue. Thank you.

[Moritz Beck]

Hi thank you both for your advice and interest in this topic. I will have a deeper look in the bias adjustments to reduce this effect. Nevertheless, it would still be interesting to me why this effect occurs.

Here is a picture of the fan I used. In this example I increased the ON threshold and decreased the OFF threshold to reduce the number of events. It's possible that the lighting emphasizes the effect of an event tail in this case.

I tried another example with a backlight and used the standard DAVIS346 parameters. This is the general scenario with a wodden ball, that appears without shadows due to the backlight.

And this is the result I get with the DAVIS346:

Thank you very much!

Moritz Beck

[Chenghan Li]

Hi Mortiz, thank you for posting the photos and providing more details. I'm afraid I still can't provide a satisfactory explanation to your observation but I'll try to share my hypothesis.

You mentioned you increased the ON threshold and decreased the OFF threshold in the rotating fan example. I suspect that the increased ON threshold caused the ON events to be much harder to trigger by the trailing edge of the fan blade. In the event output viewer, the fan blades appear much wider than they actually are, which means the trailing edge ON events are delayed much more than the leading edge ON events. If the ON threshold is set too high such that the trailing edge can barely trigger ON events, then the triggered ON events are more unevenly spread out due to noise. So my guess is that if you decrease ON threshold to a similar level as OFF threshold, you should see at least a cleaner and tighter outline of the trailing edge, which should make the fan blades appear to have the real width in the event output viewer. BTW, it's probably more informative to view the raw events by disabling the noise filter.

As for why leading edges seem cleaner than the falling edges in general, I don't think the sensor can distinguish between leading and trailing edges. I think what you observe is a result from a combination of factors, such as bias setting (OnBn, OffBn, and perhaps also RefrBp that determines how long a pixel remains in reset after triggering an event), noise filtering, the speed of the movement as well as some lighting effects.

I will try to see if I can reproduce your observed effects when I get the chance and keep you updated.

[Moritz Beck]

Thank you for your explanation! That is quite helpful for me to understand the bias configurations a bit better. I will have a deeper look into that.

It would still be great if you share similar observations or any further explanations. Maybe that could help others as well.