woodfrog / heat Goto Github PK
View Code? Open in Web Editor NEWCode for "HEAT: Holistic Edge Attention Transformer for Structured Reconstruction", CVPR 2022
Home Page: https://heat-structured-reconstruction.github.io/
License: Other
Code for "HEAT: Holistic Edge Attention Transformer for Structured Reconstruction", CVPR 2022
Home Page: https://heat-structured-reconstruction.github.io/
License: Other
I want confirmation about the warning on having Pillow and PIL in the same environment as I'm trying to perform a transfer training, and I'm receiving this message:
ImportError: cannot import name '_imaging' from 'PIL' (/home/jovyan/.local/lib/python3.8/site-packages/PIL/init.py)
after trying to run this line:
ython train.py --exp_dataset outdoor --epochs 800 --lr_drop 600 --batch_size 16 --output_dir ./checkpoints/ckpts_heat_outdoor_256_ENS --image_size 256 --max_corner_num 150 --lambda_corner 0.05 --run_validation --resume ./checkpoints/ckpts_heat_outdoor_256
I hope this email finds you well. My name is Carlos Campoverde, and I am a student at ITC-UT. I am reaching out to express my excitement about your model and the exceptional work you did on "HEAT: Holistic Edge Attention Transformer for Structured Reconstruction."
Currently, I am conducting research for my master's thesis "AUTOMATIC BUILDING ROOF PLANE STRUCTURE EXTRACTION FROM REMOTE SENSING DATA" and would like to explore the application of your model for plane roof extraction. I have conducted experiments on your dataset by following the instructions outlined in the README file. However, I am unsure about how to apply your model to new images and would appreciate your guidance.
I am writing to kindly ask for your assistance in using your model to make inferences on a new dataset of building images, each of size 256x256. Your valuable help would be greatly appreciated.
Thank you for your time and consideration.
Thanks for sharing the code!
But I'm having some problems trying to use this code.You mentioned in readme.txt that the python version used is 3.7, but after I run
$ conda create -n try python=3.7
$ conda activate try
$ pip install -r requirements.txt
I met an error:
So I changed to version 3.8. After this I compiled without problems, but when trying to run:
$ python infer.py --checkpoint_path ./checkpoints/ckpts_heat_s3d_256/checkpoint.pth --dataset s3d_floorplan --image_size 256 --viz_base ./results/viz_heat_s3d_256 --save_base ./results/npy_heat_s3d_256
I met a new error:
This looks like a version mismatch issue.
So I would like to ask about the version of CUDA, cuDNN, pytorch and cudatoolkit in the environment where you run the code, so that it can run successfully, thank you very much!
Thank you for releasing your code!
I notice that you only compared its performance on two tasks: outdoor architecture reconstruction & floorplan reconstruction.
Do you try HEAT on other datasets, like YorkUrban and Wireframe?
Hi @woodfrog thank you for the great work! I think one missing point of this codebase is that what's the intuition behind the usage of normals, which is not discussed in paper or in this codebase. What's your idea of incorporating normals into density maps as the input to HEAT? Is there any reference?
I'd also like to have you clarify some details in your codebase. Especially, for the below two snippets:
heat/s3d_preprocess/DataProcessing/PointCloudReaderPanorama.py
Lines 163 to 175 in c73a690
normals[::10]
here? I guess it's kind of a sampling operation on the dense point cloud. How did you decide this hyperparameter 10
? I personally found that a lower number such as 5
works better for my data.heat/datasets/s3d_floorplans.py
Line 51 in 66aca88
np.maximum
with the 3-dimensional (x, y, z) normals map?Hi, Thanks for your great work! I have a question regarding the ground truth of Structured3D floorplan. For quantitative evaluation of floorplan reconstruction, it is mentioned that you also used the data used by MonteFloor (downloaded from here). I downloaded the data and found the ground truth corner is actually preprocessed. Below is a comparison (scene_00030) between the GT label you used and the GT I generated from Structured3D annotations. It shows that in the GT label you used, corners and edges of adjacent rooms are merged. Could you please explain how the adjacent rooms are merged? Another thing is that it seems that the GT label has been re-annotated (see bottom right). Could you please also explain this? It would be great if you can provide the preprocessing script. Thanks in advance!
Hello, I am reproducting HEAT.
The correct one should be raise NotImplementedError('Cuda is not available')
Hello @woodfrog ,
I'm interesting with your research and i want to apply it with other datasets . Can you give me the name of tool to label or code to label ?
Thanks you so much!
How should I modify the code to run HEAT without using the data in the det_finals folder?
Hi, I find there is a postprocessing after getting the planar graph, which extracts room from the edges. But I can not find a detailed illustration of this part in the paper. Can you give an explanation of how it works?
Hi there,
Thank you very much for sharing your work. It is amazing!
I was going through the inference script and i noticed that you use normals along with the density maps and these normals are either [0,255]. However, the normals calculated during data generation are between 0 and 255. May I know what threshold (between 0 and 255) you used to get the normal maps for your inputs? It would be really nice if I get an idea here. Looking forward to your response!
Thank you very much for your open source code. I build floorplan on structured3d with very good results! Now, I run this algorithm on my own complex dataset, which has average/maximum numbers of corners 300/800(structured3d is 22/52), so the algorithm will use huge GPU memory and run failed,maybe because the initial number of edges is set to the square of the number of corners(O(N^2))? So do you have any good solutions? Looking forward to your reply!
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.