Closed chrockey closed 3 years ago
That's a great question! With ScanNet, we have tried with on BB. You should get around 66 mAP@0.25 from scratch. For SUN RGB-D, we have not tried. You are welcome to try it with yourself.
With the pretrained model from here, we used a backbone model which is about 3x larger than the original one, we are able to achieve around 69 mAP@0.25 for ScanNet and 63 mAP@0.25 for SUN RGBD.
@zaiweizhang Thank you for your kind reply!
@chrockey hi, I have the same question as you. So did you test the single backbone. If so, can you share the result. Thank you very much!
@xuntan97 I have personally tried it several times. It was 66 mAP@0.25 from scratch for scannet. And with the pretrained model from here, we used a backbone model which is about 3x larger than the original one, we are able to achieve around 69 mAP@0.25 for ScanNet and 63 mAP@0.25 for SUN RGBD. Hope this helps!
@zaiweizhang Thank you for your reply. I still have some questions. In your paper, you give the mAP 60.1% on SUNRGBD. which scale does it use? And what do you think about the args--use_color, should we set use or not use.
That's using four BB without pretraining from the method described here. You should not set the --use_color, it does not give performance gain and sometimes give worse results.
Hi,
Thank you for sharing this great work! I have a simple question about the performance with single backbone. What is the mAP@0.25 and mAP@0.5 on both ScanNet and SUN RGB-D when the model use single backbone?
Regards,