fcos3d :demo/mono_det_demo.py wrong visuliaztion results

[DMing11]

I got the result as below when run demo/mono_det_demo.py in the master branch

and run /tools/test.py got the result as below

I have tried setting several --score-thr to filter low confidence boxes, but don't got the same results as #542
need to using NMS to filter detection results?
Thanks

[Tai-Wang]

Could you please try with the threshold 0.15? There will also be overlapped boxes left because we do not apply nms among feature levels, but I think it should not be so many.

[DMing11]

Thanks for your reply. Yes I have tried with the threshold 0.15 and there are also be overlapped boxes left, but I think a detected object should eventually have only one box, like #542 , his results look beautiful , I want to know how he did it .Is it just adjusting the threshold?
thanks!

[Wuziyi616]

Hi @DMing11 I am the author of #542. The visualization with only one box per object is simply because I use GT instead of the predicted results from FCOS3D. You can see the 1st, 3rd and 4th images I post there also have many overlapped boxes for each object. I think your first image result is very similar to what I got in #542?

[DMing11]

So it is ! Thank you very much ! @Wuziyi616

[Treemann]

Hi, @DMing11 @Wuziyi616 @Tai-Wang
I met the same problem of keeping so many bboxes overlap with each other and I found it's caused by the setting of 'nms_thr'. The default setting is 'nms_thr=0.8', you could set it to a lower value (like 0.2) and the visualization result will be much better.