kaiyi98

> for training (sampling positives and negatives) or test (evaluation)? both, because my dataset is garage image, a smaller error is required.

And then，I test oxford-v1.0 dataset using your oxford-v1.0_pretrained_l5_w3 model and get the reacall@1=0.7424 recall@5=0.8775 racall@20=0.9632. But the results in paper Table I are SeqNet(S5) recall@1/5/20=0.62/0.76/0.88. I don't konw why. Thank...

I have solved it!

> You can also use the key `is_intersection_or_connector` to filter the lane segments within the intersection. but I need to keep the centerlines that the current car might travel on.

> Hi, maybe you can utilize the direction and topology of the centerlines? The order of points represents the direction of centerlines. just the vector from startpoint to endpoint？But when...

> > And then，I test oxford-v1.0 dataset using your oxford-v1.0_pretrained_l5_w3 model and get the reacall@1=0.7424 recall@5=0.8775 racall@20=0.9632. But the results in paper Table I are SeqNet(S5) recall@1/5/20=0.62/0.76/0.88. I don't konw...

> > > > > > > > > > Thanks very much for your help. According to this [rpautrat/SuperPoint#164](https://github.com/rpautrat/SuperPoint/issues/164), I comment the following three lines: > > > https://github.com/shaofengzeng/SuperPoint-Pytorch/blob/6e5c6587311cd4f98f9a5b61e84731555778c958/solver/loss.py#L126...

感谢回复，我用您的代码训练第一步，人造图形训练magicpoint，怎么到20个epoch左右验证损失就收敛了，但是原论文说magicpoint训练了200000个iteration，然后给coco打点，打出来的点效果并不理想。还望解答，谢谢。

> I think this is mainly for better performances 但是0.001的阈值会导致点特别多，做nms时间瓶颈很大，在训练superpoint的时候gpu利用率就很低，训练很慢，请问有什么方法解决的吗？我暂时是把这阈值调大了。