search

tyagi-iiitv / PointPillars

GNU General Public License v3.0
105 stars 47 forks source link

Better predictions #15

Closed nschein closed 3 years ago

nschein commented 3 years ago

This finds accurate bounding boxes on a small data set. I added an opencv requirement because the have NMS for rotated bounding boxes. Also, the background class is removed again, because class and bbox prediction work now. Note, I changed the loss weights away from the original paper, because that lead to far better results for my tests.

tyagi-iiitv commented 3 years ago

Thanks Nico!

tyagi-iiitv commented 3 years ago

Hi Nico, just wanted to know if you tested the network on the Kitti dataset? Or was it some other dataset?

nschein commented 3 years ago

Hey, I used the Kitti data set. I trained for 24k iterations on 5 samples for initial tests. Admittedly, this is a small sample size, but in previous versions this was already making troubles. I did not do larger tests so far. Counting on the community ;-)

tyagi-iiitv commented 3 years ago

sounds good, I'm actually training it on the complete Kitti dataset this time, let's see how the results turn out to be. BTW, do you remember the approx. loss value (total and occupancy) for the test you ran (0.7 occ threshold) by any chance?

nschein commented 3 years ago

Sure. I expect, however, that you won't reach quite these values -- in meaningful time - when training on all the data at the same time. angle loss: 3e-4, clf loss: 2e-5, heading loss: 8e-6, localization loss: 1e-6, occupancy loss: 1.4e-3, size loss: 2e-6 total loss: 1.4e-3