search

haoyu94 / RoITr

Rotation-Invariant Transformer for Point Cloud Matching
MIT License
108 stars 13 forks source link

About how the results deal with different downsample ratios used in different methods #2

Closed chalth closed 1 year ago

chalth commented 1 year ago

Hi, your work is excellent! However, I have a question to confirm.

In the comparison algorithm, the downsampling rate of the original 'predator' is 2.5cm, while 'geotransformer' has a 5cm downsampling rate, and the original 'lepard' has a 10cm downsampling rate.

My idea is that the final downsampling rate of these methods is different, and the predictable corresponding number is also different, resulting in a large change in the IR indicator. I guess other original methods like SpinNet and YOHO are also done at a 2.5cm downsample rate, but I haven't verified it yet.

So how did you deal with this situation in your thesis? Should all methods be unified under the 5cm downsampling rate, or should we directly compare IR under different downsampling rates without considering this factor?

In addition, the original lepard actually predicted a small number of corresponding quantities, about 1-3000, and you set the sampling number to 5000 when comparing, how did you deal with this problem?

chalth commented 1 year ago

Sorry, I got the downsampling rate of each method wrong. Predator is 0cm; Geotransfomer is 2.5cm; Lepard is 5cm.