about paper about toponet HOT 3 CLOSED

Hi, @eeeric1. Thank you for bringing this up. We apologize for any inconvenience caused.

Regarding the centerline supervision, we utilize polylines to represent centerline and feed the label to the VectorMapNet and MapTR.
For VectorMapNet, the pre-processing of their approach is remained, e.g., generate keypoints (bounding box), simplify the polyline, and apply quantization to coordinates. (for their polyline generater).
For MapTR, the centerline is interpolated at equidistant intervals to create 20 points, which serve as training label.

As for the topology reasoning, we first generate the instance queries of both approaches, based on their design in classification score prediction.
For VectorMapNet, we use the concatenate of 2 keypoint queries as the instance queries.
For MapTR, we calculate the average of the point queries to obtain the instance queries.
After that, a same topology head in TopoNet is applied to both of them to produce the driving scene topology.

from toponet.

I don't have a specific reason about this case, but here are some suggestions:

• Double-check the dimension you are averaging on. If the dimension of the query is (b, n_inst, n_pts), the average should be performed on n_pts.
• Verify the lclc/lcte label and assign the matrix correctly.
• Ensure that the order of the output sequence remains unchanged during post-processing, unless you also modify the order of the lclc/lcte matrix.
• Review the evaluation process thoroughly. You can begin with the visualization of lclc/lcte to check the results.

from toponet.

Thanks for the reply!
Regarding topology reasoning for MapTR, I first get the point queries, calculate the average of the point queries, and send the average to the topology head.
My result : The detection of lane lines and traffic lights are normal, but the topology(both lclc and lcte) score is very low. I don't know if I missed something.

from toponet.