search

skyhehe123 / SA-SSD

SA-SSD: Structure Aware Single-stage 3D Object Detection from Point Cloud (CVPR 2020)
492 stars 106 forks source link

we can achieve the 3d mid to 84.50. but the bev mid still around 90.0, what wrong with me #63

Open yangyongguang opened 4 years ago

yangyongguang commented 4 years ago

your's Car AP@0.70, 0.70, 0.70: bev AP:96.55, 92.79, 90.32 3d AP:93.13, 84.54, 81.71

mine Car AP@0.70, 0.70, 0.70: bev AP:96.51, 90.37, 87.76

we want to know, you have split the train dataset as 8:2 to train a model?

abhigoku10 commented 4 years ago

@yangyongguang are you achieving 3d ap for multiclass code or single class ??

yangyongguang commented 4 years ago

@yangyongguang are you achieving 3d ap for multiclass code or single class ?? single class 3d AP mid:84.32 is the best。 and I wand how to split the training dataset to get the BEV mid value(92.79)。

vehxianfish commented 3 years ago

single class 3d AP mid:84.32 is the best

Hi, single class and single gpu only implement 3D AP mid:83.98 on traing dataset 1:1. I have test for two months.Can you tell me your config and dataset split. Can you tell me your dataset augmentation:eg.random rotate or others? Do you use groud truth？ Thank you very much.It puzzles me a lot