StreamPETR
StreamPETR copied to clipboard
exiawsh
about customdataset
hello, i want to use my dataset to train the streampetr. what do u think about that, 500 frames is enough to let the transformer convergent or not. the performance of anchor based method is good, but i dont use query transformer based approach yet. the scenes is same in the dataset. its a fixes lidars and cameras setting on the infrastructures.
500 frames is not enough. Especially for PETR. You can try RepDetr3D, which would be better.
@exiawsh i use both point cloud and camera feature. is also not enough right? can u tell me why repdetr3d is better in small dataset size. because of deformableattention?
Yes. PETR needs more data to learn the 2D-to-3D projection. While deformable attention can do it.
@XLiedInfi I wonder know which type you tranformed your customdataset to,is it Kitti? I have not find any tools to transfer my own dataset to nuscenes, thus i processed it as kitti form. but after i changed the config profile and began to train, the print info showed grad_norm=nan and loss=0,I don't know why.If your data form is kitti and you succeeded in running,can you share me the train config and dataset config with me?I would be appreciate!
