Running in real-time on live data

[Srijal97]

Hello! I'm able to set up and run MonDi on the VOID dataset and get evaluation results similar to what is presented.

Evaluation results:
     MAE      RMSE      iMAE     iRMSE
  30.884    87.478    15.308    38.333
     +/-       +/-       +/-       +/-
  23.885    79.675    18.747    51.171
Total time: 13383.36 ms  Average time per sample: 16.73 ms

How could I go about running MonDi in real-time on live images? I noticed that the dataset images are preprocessed to be in triplets. Would this have to be done for real-time images too? Also, my depth stream is as a pointcloud; would that have to be saved as PNG depth images as in the dataset?

Thank you!

[edhyah]

@alexklwong Hey Alex, similar question - is it possible to feed it dense depth maps? I'm feeding it a dense depth map from an Azure Kinect and am struggling to see good results, but not sure if it's because it's dense or there's some other parameter I'm inputting wrong.

[alexklwong]

Hi, it looks like the code base is an old version of the original (in fact, I looked at the training loop and it looks like there is an error). I will ask the students to prepare the new code and add an online inference script.