Thanks for the video. Quite interesting that the original photo you shared showed the pillar prediction a whole extra car width away, which I suppose is better than predicting a pillar going through an imaginary car.
View attachment 1002022
The pillar disappears pretty quickly in your video, so I guess object permanence isn't that strong for these current predictions. Presumably these were trained to be most accurate for the nearest objects while further objects could be false positives based on more likely patterns from training data.