search

orlitany / DeepFunctionalMaps

90 stars 23 forks source link

How to generate .mat files? #3

Open LightRay2 opened 6 years ago

LightRay2 commented 6 years ago

Hello! I tried to reproduce .mat files for test models, but didn't succeded. 10 num bins + radius 9 are correct for SHOT? Radius 9 is applied as the percent of the area?

Best regards, Michael.

orlitany commented 6 years ago

Hi, I'll have to find the parameters i used. In the meanwhile, here's a link to the data

LightRay2 commented 6 years ago

For those who faced the same problem: .off file coordinates should be premultiplied by 100 before calculating SHOT.

A brief test runs on provided model: quite well for FAUST and SCAPE, but nearly random for TOSCA. Is it expected or possibly i'm doing something wrong (of course, no premultiplication for TOSCA, it's already in sentimeters)?

orlitany commented 6 years ago

the provided model was supposed to serve as a mock. I'll try to upload one that has been trained longer. For Tosca (dogs) we used a pretrained model (on faust) and refined it on several tosca pairs.

On Sun, May 27, 2018, 01:58 LightRay2 notifications@github.com wrote:

For those who faced the same problem: .off file coordinates should be premultiplied by 100 before calculating SHOT.

A brief test runs on provided model: quite well for FAUST and SCAPE, but nearly random for TOSCA. Is it expected or possibly i'm doing something wrong (of course, no premultiplication for TOSCA, it's already in sentimeters)?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/orlitany/DeepFunctionalMaps/issues/3#issuecomment-392293194, or mute the thread https://github.com/notifications/unsubscribe-auth/AEuHZqV8ZHjlzJGzA_5O1aAIGlX_k-phks5t2d4vgaJpZM4T3Yl8 .

GitHubforXiaoming commented 5 years ago

Hello! part_evecs: eigenvectors on source shape part_evecs_trans: transposed part_evecs with mass matrix correction model_evecs: eigenvectors on target shape model_evecs_trans: transposed model_evecs with mass matrix correction I want to obtain these value from other models. Can you explain more about it since these defines are not clear for me ?

Best regards. Xiaoming.

hellohawaii commented 5 years ago

Hello! I tried to reproduce .mat files for test models, but didn't succeded. 10 num bins + radius 9 are correct for SHOT? Radius 9 is applied as the percent of the area?

Best regards, Michael.

I am trying to compute the SHOT descriptors, too! May I ask what value I should assign to the key 'min_neighs' to get the same .mat file?

orlitany commented 5 years ago

Added a description of SHOT parameters and the calc_shot function to a Utils folder.

Adamwgoh commented 5 years ago

Hi, I noticed that the results calculated using the provided calc_shot matlab binary do not produce the same results at the ones provided. Because of that, I had not been able to reproduce the results of the sample mesh, let alone other FAUST models (using the given pre-trained model). Is it possible the sample data's SHOT is not calculated with the same way?

If it is, is there a way to provide the rest of the header and .cpp code necessary for us to compile one that is not tied to Matlab?

rFalque commented 5 years ago

@orlitany can you add the shot_descriptor.cpp and shot_descriptor.h as well?

orlitany commented 5 years ago

calc_shot.cpp is in Utils... alternatively you can write prof. Tombari as described on the project webpage: http://www.vision.deis.unibo.it/research/80-shot

orlitany commented 5 years ago

Hi, I noticed that the results calculated using the provided calc_shot matlab binary do not produce the same results at the ones provided. Because of that, I had not been able to reproduce the results of the sample mesh, let alone other FAUST models (using the given pre-trained model). Is it possible the sample data's SHOT is not calculated with the same way?

If it is, is there a way to provide the rest of the header and .cpp code necessary for us to compile one that is not tied to Matlab?

perhaps it's a scale issue. I think others have commented on this already, try scaling the models by a factor of 100

rFalque commented 5 years ago

There are some missing files related to the includes of the calc_shot.cpp

include "shot_descriptor.h"

I guess, the tools used are the same as the toolbox used there https://github.com/zorah/KernelMatching/blob/cb79b080abcb87a42ec01cac383751e342805e99/tools/shot/shot_descriptor.cpp

orlitany commented 5 years ago

There are some missing files related to the includes of the calc_shot.cpp

include "shot_descriptor.h"

I guess, the tools used are the same as the toolbox used there https://github.com/zorah/KernelMatching/blob/cb79b080abcb87a42ec01cac383751e342805e99/tools/shot/shot_descriptor.cpp

Yes, we used the same tools i that work. Thanks for pointing this out!

shixue-qing commented 3 years ago

Hello, I am a student and I have read a paper you published in 2017 called Deep Function Mapping. I want to know what the.cpp file is used for. I want to run it and find that it has no main function and no function that needs to be called inside.Also, I would like to know how Figure 7 came out?I really need the code of Figure 7, and I hope you can send it to me.Thanks a million.

shixue-qing commented 3 years ago

Hello, I need your Shrec '16 data set, I hope you can send it to me, thank you very much.

orlitany commented 3 years ago

try to follow the code here: https://github.com/OshriHalimi/unsupervised_learning_of_dense_shape_correspondence Oshri did a better job maintaining this repo and the difference between the methods is in the loss function so you could easily swap those.