Jetson deployed with semantic information not outputting correctly

[Ruihyw]

Hi!Sry to bother you again! I'm having some problems deploying on jetson ros. I created the container using dockerfile, after configuring the environment I connected the realsense D435i camera and modified the topic, color and depth to odom coordinate mapping I published the static transformation using "tf", then I started the program and no error was reported. But when I try to view /viplanner/sem_image in rviz I find that the image is all black, I use rostopic echo to see that there is content output for the topic but no matter how I move the camera the content remains the same! "success" means that /viplanner/sem_image/Compressed has been successfully published.The reported error also stopped appearing after I released the static transforms Topic has output but rviz can't display it！

[Ruihyw]

I have installed the extension but it still shows that : [Warning] Pre-trained ResNet50 models cannot be used since mask2former not found And it shows that the model dictionary is wrong like the pic above

[Ruihyw]

I found the problem, the semantic segmentation module is able to read the realsense image data, but the output is all black, I used the model and parameter files recommended in Install.md, but still get this result, I'm very confused don't know how to modify it:(

[Ruihyw]

print("Result unique values:", np.unique(result)) output: Result unique values: [133] This means that the model partitions out only one type

[pascal-roth]

Hi, This is an issue with the mask2former model, more specifically, that the downloaded weights do not fit the used architecture (I guess there was a renaming of part of the network). If I remember correctly, then this was a version issue of mmcv and mmengine, as they renamed parts of the network without updating the weights. Can you make sure that you have following versions of the packages:

mmcv 2.0.0 mmdet 3.1.0 mmengine 0.8.4

[yzysmile]

Hi, This is an issue with the mask2former model, more specifically, that the downloaded weights do not fit the used architecture (I guess there was a renaming of part of the network). If I remember correctly, then this was a version issue of mmcv and mmengine, as they renamed parts of the network without updating the weights. Can you make sure that you have following versions of the packages:

mmcv 2.0.0 mmdet 3.1.0 mmengine 0.8.4

hi, how can i get model config about image segemantation as i just see a model included its weights in ros install

[Ruihyw]

Hi, This is an issue with the mask2former model, more specifically, that the downloaded weights do not fit the used architecture (I guess there was a renaming of part of the network). If I remember correctly, then this was a version issue of mmcv and mmengine, as they renamed parts of the network without updating the weights. Can you make sure that you have following versions of the packages: mmcv 2.0.0 mmdet 3.1.0 mmengine 0.8.4

hi, how can i get model config about image segemantation as i just see a model included its weights in ros install

hi! For me,I have following versions of the packages mmcv 2.0.0 mmdet 3.1.0 mmengine 0.8.4 The model named "Swin-B" workd! You can get these data here https://github.com/open-mmlab/mmdetection/tree/master/configs/mask2former

[Ruihyw]

Hi, This is an issue with the mask2former model, more specifically, that the downloaded weights do not fit the used architecture (I guess there was a renaming of part of the network). If I remember correctly, then this was a version issue of mmcv and mmengine, as they renamed parts of the network without updating the weights. Can you make sure that you have following versions of the packages:

mmcv 2.0.0 mmdet 3.1.0 mmengine 0.8.4

Thank you so much! I changed the model,and it's workd! Now,should i use the "LIO" or "VIO" or anything else to provide the "odom"?

[pascal-roth]

That is up to you, basically the odom that is more precise.