Do not have a good result on TUM

[Zhaoyibinn]

Hello, thanks you for your open access,but I have met some problem during running.
When I use the replica dataset, there are basically no problems running, but when I use the TUM dataset, my results are significantly lower than those in the paper and are basically unusable:
PSNR: 17.38
ATE RMSE : 2.75

Is there any problem??

[AlexMorgand]

Same here.

[Lucasmogsan]

I have similar problem with even worse RMSE (>20) just running it as is on the TUM freiburg1_desk:

python -W ignore gs_icp_slam.py --dataset_path /home/GS_ICP_SLAM/dataset/TUM/rgbd_dataset_freiburg1_desk --config /home/GS_ICP_SLAM/configs/TUM/rgbd_dataset_freiburg1_desk.txt --rerun_viewer

I achieve the following:

System FPS: 29.94
ATE RMSE: 33.15
PSNR: 13.92
SSIM: 0.635
LPIPS: 0.434

... which is not good...
I use a modified docker-container on ubuntu and have a RTX 3060 Graphics card.

@bogus2000 @Riboha : Is this because some parameters are specifically tuned to particular datasets?

[nico0704]

Have you tried running it with the same parameters as provided in tum.sh?

overlapped_th=1e-3
max_correspondence_distance=0.03
knn_maxd=99999.0
trackable_opacity_th=0.09
overlapped_th2=1e-3
downsample_rate=5
keyframe_th=0.81

python -W ignore gs_icp_slam.py --dataset_path /home/GS_ICP_SLAM/dataset/TUM/rgbd_dataset_freiburg1_desk --config /home/GS_ICP_SLAM/configs/TUM/rgbd_dataset_freiburg1_desk.txt --overlapped_th=1e-3 --max_correspondence_distance=0.03 --knn_maxd=99999.0 --trackable_opacity_th=0.09 --overlapped_th2=1e-3 --downsample_rate=5 --keyframe_th=0.81 --rerun_viewer

maybe even try a higher value for the max_correspondence distance like 0.1

This worked for us. Replica might behave quite different in comparison to real-world datasets due to its synthetic nature, hence the parameters should differ as well to get good results.

[Lucasmogsan]

@nico0704 Thanks for your reply!

We did get some better results with the parameters from the tum.sh script, thank you!
Will look more into this method - have you been trying it on other or your own datasets?

[nico0704]

@Lucasmogsan

We tried it with both TUM and our own dataset and got some alright looking results.
You can check out our fork of the repo which includes adaptions for using custom datasets as well as some basic scripts to run a live version of the algorithm where images get "streamed" from a realsense camera and then processed by the algorithm.
We included some visualizations of our own data as well if you're interested :)

https://github.com/nico0704/GS_ICP_SLAM

[Lucasmogsan]

@nico0704 amazing work and thanks for sharing. I think we can make good use of your contribution indeed.

[Aoko955]

We tried it with both TUM and our own dataset and got some alright looking results. You can check out our fork of the repo which includes adaptions for using custom datasets as well as some basic scripts to run a live version of the algorithm where images get "streamed" from a realsense camera and then processed by the algorithm. We included some visualizations of our own data as well if you're interested :)

https://github.com/nico0704/GS_ICP_SLAM

@nico0704
Thank you for your good work. I am trying to use your code to run my dataset. But an error was reported with the following message：
[pcl::FastGICP::setInputTarget] Invalid or empty point cloud dataset given!
[pcl::KdTreeFLANN::setInputCloud] Cannot create a KDTree with an empty input cloud!
[pcl::FastGICP::setInputSource] Invalid or empty point cloud dataset given!
[pcl::KdTreeFLANN::setInputCloud] Cannot create a KDTree with an empty input cloud!

Is this because my dataset is bad? Can you upload your dataset to your repository?