filtering mechanism on SESS does not achieve good results

[yyuxin]

Hi, Thank you for sharing such great work! As far as I understand, 3DIoUMatch is a filtering mechanism on sess.

But I try to filter on sess by setting the threshold of score and objectness, the effect is not good, do you know what is the reason? The thresholds are selected as 0.9.

[yezhen17]

Hi @yyuxin,

We are not simply filtering upon SESS. For one thing, we do not supervise every proposal from the student model like SESS.

I don't know how you modified SESS so I have no clue. Maybe you can compare your implementation with our code?

[yyuxin]

If we don't consider 3D IoU, the difference with sess is SESS: 128 proposals from teacher network, students also have 128 proposals, do consistency loss

Then 3DIoUMatch：The 128 proposals from teacher network, after filtering, only some high quality are left, and 128 proposals from student, do loss Do I understand correctly?

[yezhen17]

Yes, this covers most of the difference except our IoU module-based operations. I think the difference between the losses is crucial to the performance gap.

[hughw19]

One key difference to blindly supervise consistency: we perform selective supervision ( see the relevant section for more details). This is very important for our performance.

On Sat, Apr 17, 2021 at 5:18 AM yyuxin @.***> wrote:

If we don't consider 3D IoU, the difference with sess is SESS: 128 proposals from teacher network, students also have 128 proposals, do consistency loss

Then 3DIoUMatch：The 128 proposals from teacher network, after filtering, only some high quality are left, and 128 proposals from student, do loss

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/THU17cyz/3DIoUMatch/issues/6#issuecomment-821813982, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEXSGH4MYQFSHBDQ3H2Z7BTTJF37ZANCNFSM43C6WD5A .

[yyuxin]

got it! Thanks for your patient reply.