Instance Segementation Input Requirement

[Kyle-Xu001]

Hi there!

When I run the MASK-RCNN to generate the preprocessed result, the label output is quite different. For example, in the first image, the background is 0, object A is 1, the object B is 2. However, in the second image, object A is 2, but object B is 1 (the background is still 0).

How to fix a problem like this to correlate the instances across different frames? Does the VDO slam system require the semantic instances to be aligned across all frames?

Thank you very much.

[jessemorris]

Hi, no VDO does not need the semantic instance to be aligned between frames. The expected output would be per instance, such as, car 1, car 2, pedestrian 1 and pedestrian 2, etc., just adding different labels to different objects of the same category. The numbers only apply in one particular frame and is NOT temporally consistent.

Sorry for slow reply