A question about cost_mat in metric

[Aguin]

Hi @yuliangguo ,
I think the implementation here https://github.com/yuliangguo/Pytorch_Generalized_3D_Lane_Detection/blob/master/tools/eval_3D_lane.py#L165 is not exactly consistent with the formula in the paper.

Besides, I also want to know why not use num_match_mat or something like match_ratio as the cost_mat.

[yuliangguo]

Hey, thanks for the questions. I am trying hard to recall the details. Hopefully the following answers your questions.

1. I think you are right, there might be a typo in the paper. The middle case might not exist. The reported result should respect the code.
2. There are some technical issues via using this MinCostFlow tool. First, it seems to only accept integer types. Second, the optimizer sometimes fails when using num_match_mat.

Ideally, the best solver should be the BSDS's bipartite matching for boundary matching, https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/. However, it is old code implemented in matlab. I chose to implement MinCostFlow as a modern alternative way in python, but it has some limitations in the flexibility.

[yuliangguo]

Hey there, thanks for the questions. I am trying hard to recall the details. Hopefully the following answers your questions. 1. I think you are right, there might be a typo in the paper. The middle case might not exist. The reported result should respect the code. 2. There are some technical issues via using this MinCostFlow tool. First, it seems to only accept integer types. Second, the optimizer sometimes fails when using num_match_mat. Ideally, the best solver should be the BSDS's bipartite matching for boundary matching, https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/. However, it is old code implemented in matlab. I chose to implement MinCostFlow as a modern alternative way in python, but it has some limitations in the flexibility.

…

On Sun, Jul 3, 2022 at 10:06 PM Guanyi Yu ***@***.***> wrote: Hi @yuliangguo <https://github.com/yuliangguo> , I think the implementation here https://github.com/yuliangguo/Pytorch_Generalized_3D_Lane_Detection/blob/master/tools/eval_3D_lane.py#L165 is not exactly consistent with the formula in the paper. [image: 2022-07-04 12-50-53 的屏幕截图] <https://user-images.githubusercontent.com/16891787/177083664-e7edc892-b814-4fc4-97f5-c910a13bf96f.png> Besides, I also want to know why not use num_match_mat or something like match_ratio as the cost_mat. — Reply to this email directly, view it on GitHub <#22>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADFMXC2GBIPR5T4FKB345ILVSJWMNANCNFSM52R4PRWQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>