Zoe-Wan
Zoe-Wan
Hello @chentingpc , does the supervised comparison model resnet50 1x use the same hyperparameter, optimizer and the scheduler as the example settings as follows? ( batch_size: 256, baselr: 0.1, weight_decay:...
i got the same problem here! And I can't find the solution, maybe I'll try conda later...
emmmm... I'm not sure, I can't find the steer angle exceeding 90, but I think... shouldn't it be between -270 and 270? And another thing that confuses me is why...
oh, I have the same trouble and don't know what to do neither!! :( Could somebody help us!!
Wow, thank you a lot!! It must be helpful (for my presentation tomorrow) no matter if it is the same. :)
I have met the very same problem when I am using WSL2 in Windows10. In my case, I suppose it is because WSL2 has not yet supported Vulkan API ,...
@xhjiang-pixel Hi!Thank you for your great research and share your work! Could I also get the nuScenes config? My email is: [[email protected]]([email protected]) Thank you!
我的理解是Object-Centric就是关注每个object在时序中的变化吧,所以用历史的object feature对做query。 BEV方法(例如videobev)只是单纯把历史帧的特征concat或者propagate到当前帧中,考虑到BEV的特征图不是每一块/像素都含有有助于3d检测的语义的,所以存在冗余信息,并且计算量比较多。 然后perspective(例如petr v2)的方法也是一样,在设计query的时候并没有考虑历史信息。它的时序部分在于考虑了ego车辆运动导致的3d坐标系偏移(也就是这篇文章里提到的MLN);以及在当前帧的query也和历史perspective feature进行交叉注意力运算。虽然可能这里的query用了某些方法使其变得稀疏(具体去看petr v2)但是由于要一直保存历史perspective的featrue(重复内容多),在长时间的情况下开销大。(虽然也不见得,因为论文中petrv2的实验好像也就考虑了两帧,但是如果考虑得更多那么确实开销很大) object centrc就是从object的角度来使用这个时序信息,即 不保存历史的perspective feature,而是保存历史object feature。使用历史的object featrue来指导当前帧query的生成。当然,每帧肯定还会有新object被识别,因此除了依历史object生成的query之外还需要和普通识别那样初始化新的query(就是论文里的initial query)。
另外,baidu在nips23上发过一篇很类似的工作,你可以看看https://proceedings.neurips.cc/paper_files/paper/2023/file/ef0dcb44a47185f5bacac62571f6e920-Paper-Conference.pdf (虽然这篇还引入了点云信息)
> 另外,baidu在nips23上发过一篇很类似的工作,你可以看看https://proceedings.neurips.cc/paper_files/paper/2023/file/ef0dcb44a47185f5bacac62571f6e920-Paper-Conference.pdf (虽然这篇还引入了点云信息) (虽然他们俩好像互相没有引用对方。。。)