|
From what I read about the rewrite it seems the first version did object labeling on a per frame base while the new version labels objects in the composite 360 overview.
The first version was fairly good in interpreting everything visible in a frame, but wasn't too good in correlating the data to other camera views and earlier frames. This hampers movement detection and depth perception, and as a consequence it puts a limit to what the neural network can make sense of.
The new version is dramatically better in that respect. Basically 2D versus 3D perception. This produces much less labeling with much higher quality.
__________________
2011 Honda Insight + HID, LEDs, tiny PV panel, extra brake pad return springs, neutral wheel alignment, 44/42 PSI (air), PHEV light (inop), tightened wheel nut.
  lifetime FE over 0.2 Gigameter or 0.13 Megamile.
For confirmation go to people just like you.
For education go to people unlike yourself.
Last edited by RedDevil; 09-25-2020 at 04:11 AM..
|