What processes in LIDAR-RCNN are specific for waymo dataset?

[QingXIA233]

Hello, just like the title saying, I wonder what are the specific processes for WOD, which means if I want to use LIDAR R-CNN on my own dataset, I have to do it differently. I already change the data_processor and everything I can think of in the loader and creat_results that are respect to waymo dataset, then I use the refined results to perform evaluation on my own dataset. However, I got NAN on rotation error, and the MAP is pretty low.

Therefore, I'm confused about some subtle processes that are performed just for waymo not for other datasets. For example, compute heading residual is necessary for using LiDAR R-CNN? Did you guys use rotation in some sublte ways? （In my dataset, the rotation is according to y axis, while in your code, it's x axis, but the way of computing rotZ is the same, I already changed it.）

This bug has been driving me crazy, that's why my issue description above is a bit messy, forgive me please. I would be grateful if you could provide me some hints. Thank you a lot. Save this almost desperate kid, please.🥺

[Lzc6996]

@QingXIA233
I am sorry for the suffering. This repo is build for the official code for our paper. So it is unfortunately all build for WOD. At least you need to adapt loader, data_utils, data_processor, config in your dataset.
The key difference may come from the processing of coordinate system. And the IOU threshold in config may should be noticed.

[QingXIA233]

Thank you for the help. Yes, I already changed the files you mentioned above. There are several questions about the code that are confusing me:

1. About the extract_points, is the heading parameter of the pred_bbox_bev important here?

Because, in WOD, the rotation is around z axis and starting from pos x axis in my understanding, but in my own dataset, the rotation is around z axis and starting form -y axis. So, if heading is used in extrac_points, should I change it according to my own dataset setting? (The heading of my dataset bboxes is shown in the figure below)

2. In loader.py, when moving gt box to pred center, why use minus here, shouldn't it just be = instead of -= ?

[Lzc6996]

@QingXIA233

1. Yes, 6 from pred_lst is heading. You should also change the extract_points.
2. this also because the definition of heading. If your dataset is different, please do some visualization to check it is correct.

[QingXIA233]

Hi, thank you for the warm patience. I just got confused by the transformations in the code:
here, when transfer pcd to pred box coords, you use the rotz matrix of proposal heading

However, when transfer gt to pred box coords, you use the rotz matrix of negative proposal heading

I didn't get the logic here, could you please enlighten me a little? Thanks

[Lzc6996]

@QingXIA233
This is the difference between left and right multiplication.

[xiuzhizheng]

@QingXIA233
Hi, what performace of lidar rcnn on your own dataset?

[QingXIA233]

@QingXIA233 Hi, what performace of lidar rcnn on your own dataset?

Hi, actually, the performance on my own dataset is surprisingly terrible. I have 11 classes, and when I tried to refine the first 4 classes, all of them got around 10 points declination. There is another issue which is about performance declination on nuscenes dataset. The author mentioned about tuning the ratio of the positive/negative. I tuned the IOU threshold and still the refinement is not acheived.

[QingXIA233]

@xiuzhizheng Did you alo tried on your own dataset? If so, how is the performance.