Open charmerfuerl opened 4 months ago
Hi, thanks for your interest in our project.
We use a different BEV coordinate system, which may be the root of the performance gap you encountered. Specifically, please check the coordinate reversion in bev encoder and the one in attention layer.
Hi, thanks for your interest in our project.
We use a different BEV coordinate system, which may be the root of the performance gap you encountered. Specifically, please check the coordinate reversion in bev encoder and the one in attention layer.
Hi, Thank you for your reply, i tried not to change the bev encoder, but also to cancel the coordinate reversion in attention layer, but still got a huge performance gap, is the reversion in bev encoder and attention layer necessary?
Other places may also be affected by the reversed y-axis, for example the temporal BEV fusion: https://github.com/yuantianyuan01/StreamMapNet/blob/main/plugin/models/mapers/StreamMapNet.py#L66.
Thank you so much! i have got the correct map. But when i changed the head embed_dims from 512 to 256, just like the maptr, the MAP decreased by 6%, so can i think it as one of the key points of the streammap method?
That's very interesting and beyond my expectation. Which config file did you use btw?
That's very interesting and beyond my expectation. Which config file did you use btw? I realized the single-frame streammap based on maptrv2, https://github.com/hustvl/MapTR. Based on the NuScenes oldsplit, the map was up to 61%, but when i changed the embed_dims from 512 to 256, without any other changes, the map was 54%.
That's very interesting and beyond my expectation. Which config file did you use btw?
I realized the single-frame streammap based on maptrv2, https://github.com/hustvl/MapTR. Based on the NuScenes oldsplit, the map was up to 61%, but when i changed the embed_dims from 512 to 256, without any other changes, the map was 54%.
That's very interesting and beyond my expectation. Which config file did you use btw?
I realized the single-frame streammap based on maptrv2, https://github.com/hustvl/MapTR. Based on the NuScenes oldsplit, the map was up to 61%, but when i changed the embed_dims from 512 to 256, without any other changes, the map was 54%.
Hello, I am also trying to implement Maptr-based streammmap recently, but I encountered some problems in the configuration process. Could you please give me some guidance?
That's very interesting and beyond my expectation. Which config file did you use btw?
I realized the single-frame streammap based on maptrv2, https://github.com/hustvl/MapTR. Based on the NuScenes oldsplit, the map was up to 61%, but when i changed the embed_dims from 512 to 256, without any other changes, the map was 54%.
Hello, I am also trying to implement Maptr-based streammmap recently, but I encountered some problems in the configuration process. Could you please give me some guidance?
Of course, how can i help you?
That's very interesting and beyond my expectation. Which config file did you use btw?
I realized the single-frame streammap based on maptrv2, https://github.com/hustvl/MapTR. Based on the NuScenes oldsplit, the map was up to 61%, but when i changed the embed_dims from 512 to 256, without any other changes, the map was 54%.
Hello, I am also trying to implement Maptr-based streammmap recently, but I encountered some problems in the configuration process. Could you please give me some guidance?
Of course, how can i help you?
hello, recently, i also releaize the streamapnet based on maptrv2, this is my wechat 15256956711, i want to ask some questions for you, thanks very much
Thanks for your great work! In terms of the single frame prediction, i think the code difference between maptr and streammapnet should be the remove of point queries and the multipoint attention, is there any other difference between them? Are the pre-process and post-process the same with maptr? I directly changed the code based on the maptrv2 code but found a huge performance degradation(almost 20% degradation).