search

christian-rauch / MultiMotionFusion

Sparse-Dense Motion Modelling and Tracking for Manipulation without Prior Object Models
https://doi.org/10.1109/LRA.2022.3200177
GNU General Public License v3.0
10 stars 2 forks source link

Reproduce result from Co-Fusion paper #3

Closed ung264 closed 2 months ago

ung264 commented 1 year ago

Hi,

Thanks for sharing the great work!

I was trying to reproduce the result of Co-Fusion and came across your repo since it supports ubuntu 20.04. I managed to build and run MMF. But when I tried to evaluate MMF on car4 and room4 sequence, I couldn't get the same result as in Co-Fusion paper, after disabling the extra command line arguments from MMF. Do you have any experience on how to reproduce CF paper result for car4 and room4 sequences? If so, could you maybe share your the arguments used? Thanks in advance.

christian-rauch commented 12 months ago

The MultiMotionFusion executable is supposed to behave like the Co-Fusion baseline in the default configuration without the proposed extensions. However, I've not tried to reproduce the results on the original klg log files.

I applied a couple of fixes on top of the upstream Co-Fusion to fix compiler warnings and errors. Those fixes can be found in the PR https://github.com/martinruenz/co-fusion/pull/20, which is the same as the fix_warnings branch in this repo. This PR/branch is sufficient, if you just want to run on the klg log format.

The interface to ROS and additional fixes are part of the rosnode branch. Some of those fixes were required to run on GPUs with less memory, e.g. I had to reduce the TEXTURE_DIMENSION from 3072 to 1024 in commit dc6dbfbaca7013ab6fc3b54d9b5d0db67cca9057.

Can you try the fix_warnings and rosnode branches and report back if they work for you?

ung264 commented 12 months ago

Thanks for sharing your experience.

When I use the fix_warnings branch, I got the same result as for master branch. I evaluate both car4-full and room4 in the format of klg and extracted RGBD images, using scripts from dataset-tools . I obtained 5cm camera AT RMSE for car4 and 14cm camera AT RMSE for room4. They are much higher than the value indicated in the CF paper. I will open an issue about this problem in the original CF repository.

christian-rauch commented 12 months ago

Before opening an issue in the Co-Fusion repo, I would suggest that you actually try to reproduce the results with the original implementation on the original klg log files.