Closed JurinJC closed 10 months ago
I have similar question that what folders should I build for step2, and where should I put the over-seg results in? I'm confused about the readme command "--dataset custom_dataset --ROOT_PATH $CUSTOM_SET_DIR", i didn't get any "custom**" named files in step1...
You need to create two folders to store the results of step 1: ../datasets/custom_set/data/
and ../datasets/custom_set/pssnet_graphs/
. The remaining folders will be created automatically. Please refer to the above reply for more information on where to place the over-segmentation results. You can ignore "custom_dataset" as we don't really use it at this stage to differentiate between other datasets.
The training data for step 2 can be generated in step 1, and the size of the data can be very large.
Hi, thanks for pointing out this issue. I have fixed this small bug in the "dev" branch and created a PR. You can update the project and use 'dev' to get the training data for step2.
if I only want to run my predict data using your model of step2, can I only run python partition/pssnet_visualize.py --dataset custom_dataset --ROOT_PATH $CUSTOM_SET_DIR --res_file "../datasets/custom_set/results/predictions_test" --output_type r? Also what is the res_file(predictions_test.h5), I don't have this file.
When I run the test data like the example, I got an error like this:
I can't fix it. Is this error related to the torch version? What is your torch version
Hi, I am using Pytorch 1.12 if I remember correctly, but I don't think this is related to the Pytorch version. The parameters of the training model don't seem to match the current code. I'm not sure what is causing this, maybe it's because of the new data. Can you try to retrain the model or test it on the SUM dataset?
Hi, actually I use the SUM dataset you provided
they are generated from step1
When I retrain the model using the SUM dataset, I got an error again, like this:
When I load the .h5 files, I cant find where the parameter "entry" comes from, what " G.vs[s]['v']" actually is.
In this code:
cloud, add_feas = load_superpoint(args, db_path + '/features/' + fname + '.h5', G.vs[s]['v'], train, test_seed_offset, s_label)
yours is "/parsed/", I dont have this folder, so I change it to '/features/'.
Please do not change any of the names of the folders as this will cause unexpected errors. You do not have "/parsed/" because you did not run the preprocessing step to generate them. You need to run the pssnet_custom_dataset.py to generate the "/parsed/" folder before you run the pssnet_main.py. Don't forget to run pssnet_partition.py to generate '/features/' at first. So the pipeline is pssnet_partition.py -> pssnet_custom_dataset.py -> pssnet_main.py -> pssnet_visualize.py.
When I run only predict data with operating_mode = PSSNet_pipeline_for_GCN, I can only get result like this
this is the classifaction
this is mesh_seg in segments
I can't run PSSNet_oversegmentation and subsequent operations, so I cant get the predict dataset for step2.
Have you updated your code and switched to the dev
branch? The current main
branch contains bugs that will be merged by dev
after the updated code has been reviewed.
I have switched to dev branch. However, when I load trained model, I always got something error.
Could you provide more information about the error?
Thank for your help, I have done the hole code, but my result is very bad, can you give my some help? (used my own data)
This is the input mesh:
This is the result of step1:
This is the seg result:
Hi, it seems your mesh contains topological errors such as duplicated or non-manifold vertices, which could lead to bad results. These errors can break the continuity of the mesh surface and can be difficult to fix, especially if the vertices are related to texture coordinates.
I have fix the input, but the result is bad too, can you give me some advice?
For me, this result is better than your previous one. You can further fine tune the "balance parameter" to get better results on your dataset. However, from your second image, the classification of planar and non-planar is not as good as expected. I guess you are not using colour information?
Thanks for your answer. Actually I have used the texture, so I think I also used the colour information.
If you use our pre-trained model on your data, the results for planar and non-planar classification are acceptable because of the significant domain gap between your data and our benchmark dataset.
Hi author: In step2, if I must train the network? Does all the inputs of step2 generated from step1(train test and predictions)?
Yours! Best regards! Jurin