HorizonNet

HorizonNet

This is the implementation of our CVPR’19 " HorizonNet: Learning Room Layout with 1D Representation and Pano Stretch Data Augmentation" (project page).

Update

• 2021.11.04: Report results on Zillow Indoor dataset. (See the report :clipboard: on ZInD).
• 2021.04.03: Check out our new project HoHoNet on this task and more!
• 2021.03.14: (1) Use mesh instead of point cloud as layout viewer. (2) Update lsd detector dependency.
• 2019.08.19: Report results on Structured3D dataset. (See the report :clipboard: on ST3D).
• 2019.06.15: Bug fix for general layout (dataset.py, inference.py and misc/post_proc.py)

Feature

This repo is a pure python implementation that you can:

• Inference on your images to get cuboid or general shaped room layout
• 3D layout viewer
• Correct rotation pose to ensure manhattan alignment
• Pano stretch augmentation copy and paste to apply on your own task
• Quantitative evaluatation of 2D IoU, 3D IoU, Corner Error, Pixel Error of cuboid/general shape
• Your own dataset preparation and training

Method overview

Installation

Pytorch installation is machine dependent, please install the correct version for your machine. The tested version is pytorch 1.8.1 with python 3.7.6.

Dependencies (click to expand)

• numpy
• scipy
• sklearn
• Pillow
• tqdm
• tensorboardX
• opencv-python>=3.1 (for pre-processing)
• pylsd-nova
• open3d>=0.7 (for layout 3D viewer)
• shapely

Download

Dataset

• PanoContext/Stanford2D3D Dataset
  • Download preprocessed pano/s2d3d for training/validation/testing
    • Put all of them under data directory so you should get:

      HorizonNet/
      ├──data/
      |  ├──layoutnet_dataset/
      |  |  |--finetune_general/
      |  |  |--test/
      |  |  |--train/
      |  |  |--valid/
      

    • test, train, valid are processed from LayoutNet’s cuboid dataset.
    • finetune_general is re-annotated by us from train and valid. It contains 65 general shaped rooms.
• Structured3D Dataset
  • See the tutorial to prepare training/validation/testing for HorizonNet.
• Zillow Indoor Dataset
  • See the tutorial to prepare training/validation/testing for HorizonNet.

Pretrained Models

Please find the pre-trained weights here. By downloading the weights, you agree with the license and term of usage of the training datasets.

• resnet50_rnn__panos2d3d.pth
  • Trained on PanoContext/Stanford2d3d 817 pano images.
• resnet50_rnn__st3d.pth
  • Trained on Structured3D 18362 pano images
  • Data setup: original furniture and lighting.
• resnet50_rnn__mp3d.pth
  • Trained on Matterport3D datasets.
• resnet50_rnn__zind.pth
  • Trained on Zillow Indoor 20077 pano images.
  • Data setup: layout_visible, is_primary, is_inside, is_ceiling_flat.

CLI downloading example:

huggingface-cli download sunset1995/HorizonNet resnet50_rnn__panos2d3d.pth --local-dir ./ckpt

Python script downloading example:

from huggingface_hub import hf_hub_download

file_path = hf_hub_download(
    repo_id="sunset1995/HorizonNet", 
    filename="resnet50_rnn__panos2d3d.pth", 
    local_dir="./ckpt")

Inference on your images

In below explaination, I will use assets/demo.png for example.

• (modified from PanoContext dataset)

1. Pre-processing (Align camera rotation pose)

• Execution: Pre-process the above assets/demo.png by firing below command.

  python preprocess.py --img_glob assets/demo.png --output_dir assets/preprocessed/
  

  • --img_glob telling the path to your 360 room image(s).
    • support shell-style wildcards with quote (e.g. "my_fasinated_img_dir/*png").
  • --output_dir telling the path to the directory for dumping the results.
  • See python preprocess.py -h for more detailed script usage help.
• Outputs: Under the given --output_dir, you will get results like below and prefix with source image basename.
  • The aligned rgb images [SOURCE BASENAME]_aligned_rgb.png and line segments images [SOURCE BASENAME]_aligned_line.png

    • demo_aligned_rgb.png	demo_aligned_line.png

  • The detected vanishing points [SOURCE BASENAME]_VP.txt (Here demo_VP.txt)

    -0.002278 -0.500449 0.865763
    0.000895 0.865764 0.500452
    0.999999 -0.001137 0.000178
    

2. Estimating layout with HorizonNet

• Execution: Predict the layout from above aligned image and line segments by firing below command.

  python inference.py --pth ckpt/resnet50_rnn__mp3d.pth --img_glob assets/preprocessed/demo_aligned_rgb.png --output_dir assets/inferenced --visualize
  

  • --pth path to the trained model.
  • --img_glob path to the preprocessed image.
  • --output_dir path to the directory to dump results.
  • --visualize optinoal for visualizing model raw outputs.
  • --force_cuboid add this option if you want to estimate cuboid layout (4 walls).
• Outputs: You will get results like below and prefix with source image basename.
  • The 1d representation are visualized under file name [SOURCE BASENAME].raw.png
  • The extracted corners of the layout [SOURCE BASENAME].json

    {"z0": 50.0, "z1": -59.03114700317383, "uv": [[0.029913906008005142, 0.2996523082256317], [0.029913906008005142, 0.7240479588508606], [0.015625, 0.3819984495639801], [0.015625, 0.6348703503608704], [0.056027885526418686, 0.3881891965866089], [0.056027885526418686, 0.6278984546661377], [0.4480381906032562, 0.3970482349395752], [0.4480381906032562, 0.6178648471832275], [0.5995567440986633, 0.41122356057167053], [0.5995567440986633, 0.601679801940918], [0.8094607591629028, 0.36505699157714844], [0.8094607591629028, 0.6537724137306213], [0.8815288543701172, 0.2661873996257782], [0.8815288543701172, 0.7582473754882812], [0.9189453125, 0.31678876280784607], [0.9189453125, 0.7060701847076416]]}
    

3. Layout 3D Viewer

• Execution: Visualizing the predicted layout in 3D using points cloud.

  python layout_viewer.py --img assets/preprocessed/demo_aligned_rgb.png --layout assets/inferenced/demo_aligned_rgb.json --ignore_ceiling
  

  • --img path to preprocessed image
  • --layout path to the json output from inference.py
  • --ignore_ceiling prevent showing ceiling
  • See python layout_viewer.py -h for usage help.
• Outputs: In the window, you can use mouse and scroll wheel to change the viewport
  • 

Your own dataset

See tutorial on how to prepare it.

Training

To train on a dataset, see python train.py -h for detailed options explaination.
Example:

python train.py --id resnet50_rnn

• Important arguments:
  • --id required. experiment id to name checkpoints and logs
  • --ckpt folder to output checkpoints (default: ./ckpt)
  • --logs folder to logging (default: ./logs)
  • --pth finetune mode if given. path to load saved checkpoint.
  • --backbone backbone of the network (default: resnet50)
    • other options: {resnet18,resnet34,resnet50,resnet101,resnet152,resnext50_32x4d,resnext101_32x8d,densenet121,densenet169,densenet161,densenet201}
  • --no_rnn whether to remove rnn (default: False)
  • --train_root_dir root directory to training dataset. (default: data/layoutnet_dataset/train)
  • --valid_root_dir root directory to validation dataset. (default: data/layoutnet_dataset/valid/)
    • If giveng, the epoch with best 3DIoU on validation set will be saved as {ckpt}/{id}/best_valid.pth
  • --batch_size_train training mini-batch size (default: 4)
  • --epochs epochs to train (default: 300)
  • --lr learning rate (default: 0.0001)
  • --device set CUDA enabled device using device id (not to be used if multi_gpu is used)
  • --multi_gpu enable parallel computing on all available GPUs

Quantitative Evaluation - Cuboid Layout

To evaluate on PanoContext/Stanford2d3d dataset, first running the cuboid trained model for all testing images:

python inference.py --pth ckpt/resnet50_rnn__panos2d3d.pth --img_glob "data/layoutnet_dataset/test/img/*" --output_dir output/panos2d3d/resnet50_rnn/ --force_cuboid

• --img_glob shell-style wildcards for all testing images.
• --output_dir path to the directory to dump results.
• --force_cuboid enfoce output cuboid layout (4 walls) or the PE and CE can’t be evaluated.

To get the quantitative result:

python eval_cuboid.py --dt_glob "output/panos2d3d/resnet50_rnn/*json" --gt_glob "data/layoutnet_dataset/test/label_cor/*txt"

• --dt_glob shell-style wildcards for all the model estimation.
• --gt_glob shell-style wildcards for all the ground truth.

If you want to:

• just evaluate PanoContext only python eval_cuboid.py --dt_glob "output/panos2d3d/resnet50_rnn/*json" --gt_glob "data/layoutnet_dataset/test/label_cor/pano*txt"
• just evaluate Stanford2d3d only python eval_cuboid.py --dt_glob "output/panos2d3d/resnet50_rnn/*json" --gt_glob "data/layoutnet_dataset/test/label_cor/camera*txt"

:clipboard: The quantitative result for the released resnet50_rnn__panos2d3d.pth is shown below:

Testing Dataset	3D IoU(%)	Corner error(%)	Pixel error(%)
PanoContext	83.39	0.76	2.13
Stanford2D3D	84.09	0.63	2.06
All	83.87	0.67	2.08

Quantitative Evaluation - General Layout

• See the report :clipboard: on ST3D for more detail.
• See the report :clipboard: on MP3D for more detail.

TODO

• Faster pre-processing script (top-fron alignment) (maybe cython implementation or fernandez2018layouts)

Acknowledgement

• Credit of this repo is shared with ChiWeiHsiao.
• Thanks limchaos for the suggestion about the potential boost by fixing the non-expected behaviour of Pytorch dataloader. (See Issue#4)

Citation

@inproceedings{SunHSC19,
  author    = {Cheng Sun and
               Chi{-}Wei Hsiao and
               Min Sun and
               Hwann{-}Tzong Chen},
  title     = {HorizonNet: Learning Room Layout With 1D Representation and Pano Stretch
               Data Augmentation},
  booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition, {CVPR}
               2019, Long Beach, CA, USA, June 16-20, 2019},
  pages     = {1047--1056},
  year      = {2019},
}