Object transformed by shadow

Knowledge Hub
1

Hello,

I open this topic about the impact of object shadow on data. I have an example that comes up every time the sun is shining.

The bikes turn into a car as soon as the sun casts a shadow on the bikes.

This Monday, April 17: overcast weather and normal traffic

Between 8:00 and 9:00

cabanis_17_avril
cabanis_17_avril780×534 37.1 KB

cabanis_17_avril_photo
cabanis_17_avril_photo480×640 112 KB

This Wednesday, April 19: sunny weather and almost all the bikes are transformed into cars.

Between 8:00 and 9:00

cabanis_19_avril
cabanis_19_avril745×552 20.9 KB

cabanis_19_avril_photo
cabanis_19_avril_photo480×640 126 KB

B > A direction is forbidden for car, but when it’s sunny, there are a lot of cars…!

I installed polarizing filters but it mostly works for “poor quality” in bright sunshine.

I also installed the reflective boxes.

So do you have methods for:

Limit this effect? (a filter or something?)

Take this anomaly into account in the data analysis?

  • Crossing the data with the weather would be an interesting idea, it seems to me that Telraam crossed this information at the start of the project.
  • Set an alert when many cars are driving in one forbidden direction (transformation of cyclists into cars)

Does the Telraam S2 make the difference between object and shadow?

Thanks for your feedback !

2

Yes, I’m afraid that such strong shading and contrast is a potential issue for the way that a V1 works (bounding boxes).

It is not a light effect that can be corrected with filters as it is about the shape of the “object” that will be detected by the pixelated image.

The S2 works in a different way and “recognises” an object as either a bike or a car. If it is a bike with a shadow, it might be the size of a car, but won’t look like one so should not be classified as one. However, it is possible that the bike will also not “look like a bike” to the machine either, so in this case it would most likely be rejected and not counted. We’d need to test it to take a look.

In terms of the data validity, we are hoping to bring data annotation to the dashboard in the near future, but there is still other priority work to be done on the S2 including night counting, so I’m not sure when it might happen.

If anyone has ideas for this (other than moving the location) I’d be interested to hear them.