search

xingxuejun1989 / SCVC

8 stars 1 forks source link

RR result of FPFH on 3DLoMatch datasets. #4

Open pppyj-m opened 2 months ago

pppyj-m commented 2 months ago

In GitHub, the RR results for FPFH and FCGF are ~62% and ~64% respectively, which is very close. However, in other papers, it indicates the FPFH performs much worse than the FCGF descriptor, like ~40% vs ~60%. Even in your paper, the FPFH is much worse than the FCGF (especially on the low overlap condition) according to the following fig.

image

So, I wonder if there exists any wrong with the FPFH RR result (~62%) in GitHub?

xingxuejun1989 commented 1 month ago

Due to the passage of time, the specific experimental details of this image have become somewhat unclear, and the experimental code for this image was cleaned up during code compilation. I roughly remember that cosine distance was used to determine the correspondence, and a specific number of correspondences were selected based on the NNDR score. Additionally, RANSAC often achieves good results when the parameters are set appropriately.

xingxuejun1989 commented 1 month ago

You can verify the results on the dataset using the provided data, code, and parameter settings, and there should not be much difference.