Open Rkyzzy opened 1 year ago
LegendBC
commented
1 year ago @Rkyzzy Thanks for your interest! VectorMapNet does not use [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]. The point_cloud_range hyperparameter is just a placeholder. It does not pass point_cloud_range to any function. It just hardcoded the range here and here, which is the same as ours. So the comparison is fair.
Rkyzzy
commented
1 year ago @LegendBC Thank you very much for the reply. After carefully check, it is as what you said. However, I still want to ask whether is there any suggestion on improving the performance or direction to improve in a larger point cloud range? (Like I mentioned [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0] ) , now it has mAP 30, I don't expect it to reach 50+, but I think at least 40+ is reasonable result. Thanks in advance!
jasmine-97
commented
11 months ago Hello , i would like to know whether you change other parameters in the config or model , when you switch the point_cloud_range to [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0]? Looking forward to your reply！
guzuoxu
commented
9 months ago @LegendBC Thank you very much for the reply. After carefully check, it is as what you said. However, I still want to ask whether is there any suggestion on improving the performance or direction to improve in a larger point cloud range? (Like I mentioned [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0] ) , now it has mAP 30, I don't expect it to reach 50+, but I think at least 40+ is reasonable result. Thanks in advance!
Hi, may I ask if your experiment in the long-range domain has been successful?
Thank you author for the great work. I have a question about point cloud range, to make a fair comparison with VectorMapNet, I switch the point cloud range to [-51.2, -51.2, -5.0, 51.2, 51.2, 3.0] (which is the point cloud range of VectorMapNet) , as the point cloud range increase, the mAP drops to ~25, which is much lower than the desired result. I also checked issues like #32 #54 , but it seems that it is still an open problem. I've tried increasing vec_num from 50 to 100 as #32 suggested, it has some increase in mAP, which makes it around ~29 , but still not desired result. I've also conduct some experiments with pts_num, num_layers, but none of them can make mAP higher. In my opinion, when perception range increases, the task become more complex, it is foreseeable to have some accuracy drop. But I'm not sure whether ~29 is the bottle neck of this range. Have you tried different range before? Or do you have any information on the direction to improve when using a larger point cloud range (maybe some important hyperparameters to be modified). Thanks in advance.