search

ISUE / relocforests

C++ implementation of Scene Coordinate Regression Forests for Camera Relocalization.
MIT License
45 stars 21 forks source link

test image #11

Open LiliMeng opened 8 years ago

LiliMeng commented 8 years ago

Hi, Thanks a lot for the code and your helpful answer!

Is the test image used for training in your code? It seems that one image ( Eigen::Affine3d pose = forest->Test(data->GetRGBImage(200), data->GetDepthImage(200)); )from all the data was used for testing, while all the data was used for training (forest = new Forest(data, settings);) the regression forest. Therefore, the result ( "found pose:" ) is training result. 

Thanks again!
connerbrooks commented 8 years ago

Yes that is correct, you are testing with a training image. That is a holdover from early development and should be resolved in the latest commits to master. I have added clarification for the example applications interface and fixed issues with the command line interface. There is now a better distinction between train and test time. Let me know if these changes make it more clear.

Test time statistics is still something I have to do.

LiliMeng commented 8 years ago

Thanks a lot! Have you tested the 7 scenes dataset? I cannot get the 31% accuracy for Chess before the pose refinement mentioned in the 4.4 Main Results part of the paper either from your code and my code.

Thanks again!

connerbrooks commented 8 years ago

I haven't tested with any of the 7 scenes data sets. What kind of accuracy are you getting?

LiliMeng commented 8 years ago

Currently almost no data could be within 5 degree rotation and 5cm translation. I'll let you know the final result as soon as I complete all the test:)

connerbrooks commented 8 years ago

I think the issue with the 7-scenes may be due to an improper camera calibration. I currently hard code the camera calibration for the TUM datasets. Implementing #6 may make this more clear.

LiliMeng commented 8 years ago

Thanks a lot! That may make sense! As the 7 Scenes data set(http://research.microsoft.com/en-us/projects/7-scenes/) use the following default intrinsics: Principle point (320,240), Focal length (585,585). It's very different from the TUM dataset: Principle point (319.5, 239.5), focal length (525.0, 525.0)