question about rosbag

[wo712268]

Hello! Thank you for your work!

I feel confused when i use the rosbag, the msg lists are:

/gt_pose_wc
/lanes_gt
/lanes_predict

I print the /lanes_predict and noticed that the points in each detected lane are sparse. The distance between first few points is 5 meters, and the rear is 10 meters, or even 20 meters.

I know little about Persformer (っ °Д °;)っ, are these points sampled? Can i test my data directly with Persformer's results?

[qiaozhijian]

Thanks for your interest.

1. Yes, points are sampled in Persformer. Specifically, Persformer will sample points along with Y axis at a fixed distance and determine whether the sampled point is valid based on the predicted visibility, see resample_laneline_in_y_with_vis. Maybe most points are predicted unvisiable in these cases.
2. You can't directly use results from Persformer for two reasons:
   • you should transform the coordinate system, referring to transform_persformer_to_openlane
   • you need odometry between two frames, which can be obtained by VIO or GT poses.

[wo712268]

Thanks for your interest.

1. Yes, points are sampled in Persformer. Specifically, Persformer will sample points along with Y axis at a fixed distance and determine whether the sampled point is valid based on the predicted visibility, see resample_laneline_in_y_with_vis. Maybe most points are predicted unvisiable in these cases.

2. You can't directly use results from Persformer for two reasons:

   • you should transform the coordinate system, referring to transform_persformer_to_openlane
   • you need odometry between two frames, which can be obtained by VIO or GT poses.

Thank you for your reply! I have odometry pose which is from VINS, thank you! (@^0^@)/

[bitwangdan]

@qiaozhijian Hi, “transform_persformer_to_openlane” , is it mapped from camera to ego?

[qiaozhijian]

@qiaozhijian Hi, “transform_persformer_to_openlane” , is it mapped from camera to ego?

Hi, sorry for late reply.

I'm sorry whether you can give the exact link to this code?

I assume you are confused about the prediction of PersfFormer. According to this issue, Persformer predicts xyz in the LaneNet vehicle-coord system, while the openlane GT is annotated in the camera system. So we need this transformation.