For real-time object detection purposes

[gnefihs]

Would this be suitable for real-time object detection, say for autonomous vehicle purposes?

My concern is that when a new object appears, will it only be detected when the next keyframe is processed? If so then, there will be a delay before an object is detected, which is undesirable for real-time object detection purposes.

If not, does the model handle the appearance of new objects between keyframes somehow?