Weird inference results

[ljwpk]

Hi, I'm trying to run the pre-trained SalsaNext model with point clouds.

I created the environment with an anaconda. All datasets for point clouds and lidar poses were downloaded and unzip.

Could you check my setup?

And then, I followed the readme file.

1. conda env create -f salsanext_cuda10.yml --name salsanext
2. conda activate salsanext
3. export CUDA_VISIBLE_DEVICES="1"
4. cd /path/to/benchmarks/SalsaNext/train/tasks/semantic
5. python infer2.py -d ~/Rellis -l ./prediction -s test -m ~/Rellis_pretrained

But outputs are weird. I attached the capture for the correlation matrix.

Thanks!

[maskjp]

Hi, @ljwpk ,

Thank you for your interest in RELLIS3D. Your setup looks good. Could you give me more information? For example, the output of cell 35 in the Jupyter notebook, or share your prediction with me.

Thanks!

[ljwpk]

Thank you for your reply. This is the Jupyter notebook file and my prediction file.

Jupyter notebook: https://www.dropbox.com/s/y1tv5zgmwac4aeu/Untitled.ipynb?dl=0 Prediction result: https://www.dropbox.com/s/5l648jet38au0bd/prediction.zip?dl=0

Capture for jupyter notebook.

[maskjp]

Hi, @ljwpk

Thanks for your reply. I checked the prediction files. Most of the predictions are 3. That's weird.

I saw you used salsanext_cuda10.yml. Could you try to create the python environment using this file?

Did you download the latest data from the github?

Thanks!

[ljwpk]

I downloaded it a few days ago.

I will try to create a new environment you gave.

LiDAR with Color Annotation PLY Format (Download 26GB) - Rellis_3D_os1_cloud_node_color_ply.zip LiDAR SemanticKITTI Format (Download 14GB) - Rellis_3D_os1_cloud_node_kitti_bin.zip LiDAR Annotation SemanticKITTI Format (Download 174MB) - Rellis_3D_os1_cloud_node_semantickitti_label_id_20210614.zip LiDAR Scan Poses files (Download 174MB) - Rellis_3D_lidar_poses_20210614.zip LiDAR Split File (75KB) - Rellis_3D_lidar_split.zip

Camera to LiDAR (Download 3KB) - Rellis_3D_cam2lidar_20210224.zip Calibration Raw Data (Download 774MB)

Thanks!

[maskjp]

Hi, @ljwpk,

Did you solve your problem?

[ljwpk]

Not yet. I could not solve the problem. Any guesses on this issue?

[maskjp]

@ljwpk ,

Can you send me a part of the lidar data you are using? Maybe the smallest sequence

[ljwpk]

@maskjp I solved this issue. There was a related library version issue.

I tried to test different OS(Windows, Ubuntu).

1. Windows My laptop consists of GTX1650 with CUDA 10.1(reqirements.txt crashed for PyTorch).

1) I remove PyTorch and torchvision in reqirements.txt 2) Install requirements. pip install -r reqirements.txt 3) Install PyTorch. conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch

2. Ubuntu Machine learning computers had two RTX 3090 with CUDA 11.2. When I tried to install it, I got the same issue. 1) & 2) are same. 3) conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Thank you, maskjp.

[maskjp]

Great! I am glad you solved your problem and thank you for providing this important information to us!

Thanks!

Best wishes!