planercnn

by NVlabs

NVlabs / planercnn

PlaneRCNN detects and reconstructs piece-wise planar surfaces from a single RGB image

249 Stars 65 Forks Last release: Not found Other 21 Commits 0 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:

License CC BY-NC-SA 4.0 Python 3.7

PlaneRCNN: 3D Plane Detection and Reconstruction from a Single Image

alt text

By Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz

This paper will be presented (Oral) in IEEE CVPR 2019.

Introduction

This paper proposes a deep neural architecture, PlaneR-CNN, that detects arbitrary number of planes, and reconstructs piecewise planar surfaces from a single RGB image. For more details, please refer to our paper and video, or visit project website. The code is implemented using PyTorch.

Project members

License

Copyright (c) 2018 NVIDIA Corp. All Rights Reserved. This work is licensed under the Creative Commons Attribution NonCommercial ShareAlike 4.0 License.

Getting Started

Clone repository:

git clone https://github.com/NVlabs/planercnn.git

Please use Python 3. Create an Anaconda environment and install the dependencies:

conda create --name planercnn
conda activate planercnn
conda install -y pytorch=0.4.1
conda install pip
pip install -r requirements.txt
Equivalently, you can use Python virtual environment to manage the dependencies:
pip install virtualenv
python -m virtualenv planercnn
source planercnn/bin/activate
pip install -r requirements.txt
Now, we compile nms and roialign as explained in the installation section of pytorch-mask-rcnn. To be specific, you can build these two functions using the following commands with the right
--arch
option:

| GPU | arch | |-------------------------|-------| | TitanX | sm52 | | GTX 960M | sm50 | | GTX 1070 | sm61 | | GTX 1080 (Ti), Titan XP | sm61 |

More details of the compute capability are shown in NVIDIA

cd nms/src/cuda/
nvcc -c -o nms_kernel.cu.o nms_kernel.cu -x cu -Xcompiler -fPIC -arch=[arch]
cd ../../
python build.py
cd ../


cd roialign/roi_align/src/cuda/ nvcc -c -o crop_and_resize_kernel.cu.o crop_and_resize_kernel.cu -x cu -Xcompiler -fPIC -arch=[arch] cd ../../ python build.py cd ../../

Please note that, the Mask R-CNN backbone does not support cuda10.0 and gcc versions higher than 7. If you have troubles compiling these two libraries, try to downgrade PyTorch to 0.4.0 before compilation and upgrade back to 0.4.1 after compilation. You might also want to find more information on their original repository.

Models

Models are saved under checkpoint/. You can download our trained model from here, and put it under checkpoint/ if you want to fine-tune it or run inferences.

Plane representation

In this project, plane parameters are of absolute scale (in terms of meters). Each plane has three parameters, which equal to planenormal * planeoffset. Suppose planenormal is (a, b, c) and planeoffset is d, every point (X, Y, Z) on the plane satisfies, aX + bY + cZ = d. Then plane parameters are (a, b, c)*d. Since plane normal is a unit vector, we can extract planenormal and planeoffset from their multiplication.

Run the inference code with an example

python evaluate.py --methods=f --suffix=warping_refine --dataset=inference --customDataFolder=example_images

Results are saved under "test/inference/". Besides visualizations, plane parameters (#planes x 3) are saved in "*planeparameters0.npy" and plane masks (#planes x 480 x 640) are saved in "*planemasks0.npy".

Run the inference code with custom data

Please put your images (.png or .jpg files), and camera intrinsics under a folder ($YOURIMAGEFOLDER). The camera parameters should be put under a .txt file with 6 values (fx, fy, cx, cy, imagewidth, imageheight) separately by a space. If the camera intrinsics is the same for all images, please put these parameters in camera.txt. Otherwise, please add a separate intrinsics file for each image, and name it the same with the image (changing the file extension to .txt). And then run:

bash
python evaluate.py --methods=f --suffix=warping_refine --dataset=inference --customDataFolder=$YOUR_IMAGE_FOLDER

Training

Training data preparation

Please first download the ScanNet dataset (v2), unzip it to "$ROOT_FOLDER/scans/", and extract image frames from the .sens file using the official reader.

Then download our plane annotation from here, and merge the "scans/" folder with "$ROOTFOLDER/scans/". (If you prefer other locations, please change the paths in *datasets/scannetscene.py*.)

After the above steps, ground truth plane annotations are stored under "$ROOT_FOLDER/scans/scene/annotation/". Among the annotations, *planes.npy stores the plane parameters which are represented in the global frame. Plane segmentation for each image view is stored under segmentation/.

To generate such training data on your own, please refer to data_prep/parse.py. Please refer to the README under data_prep/ for compilation.

Besides scene-specific annotation under each scene folder, please download global metadata from here, and unzip it to "$ROOTFOLDER". Metadata includes the normal anchors (anchorplanesN.npy) and invalid image indices caused by tracking issues (invalidindices_*.txt).

Training with custom data

To train on custom data, you need a list of planes, where each plane is represented using three parameters (as explained above) and a 2D binary mask. In our implementation, we use one 2D segmentation map where pixels with value i belong to the ith plane in the list. The easiest way is to replace the ScanNetScene class with something interacts with your custom data. Note that, the planeinfo, which stores some semantic information and global plane index in the scene, is not used in this project. The code is misleading as global plane indices are read from planeinfo here, but they are used only for debugging purposes.

Training script

python train_planercnn.py --restore=2 --suffix=warping_refine

options: ```bash --restore: - 0: training from scratch (not tested) - 1 (default): resume training from saved checkpoint - 2: training from pre-trained mask-rcnn model

--suffix (the below arguments can be concatenated): - '': training the basic version - 'warping': with the warping loss - 'refine': with the refinement network - 'refineonly': train only the refinement work - 'warpingrefine_after': add the warping loss after the refinement network instead of appending both independently

--anchorType: - 'normal' (default): regress normal using 7 anchors - 'normal[k]' (e.g., normal5): regress normal using k anchors, normal0 will regress normal directly without anchors - 'joint': regress final plane parameters directly instead of predicting normals and depthmap separately ```

Temporary results are written under test/ for debugging purposes.

Evaluation

To evaluate the performance against existing methods, please run:

bash
python evaluate.py --methods=f --suffix=warping_refine
Options:
bash
--methods:
- f: evaluate PlaneRCNN (use --suffix and --anchorType to specify configuration as explained above)
- p: evaluate PlaneNet
- e: evaluate PlaneRecover
- t: evaluate MWS (--suffix=gt for MWS-G)
Statistics are printed in terminal and saved in logs/global.txt for later analysis.

Note that PlaneNet and PlaneRecover are under the MIT license.

To evaluate on the NYU Depth dataset, please first download the labeled dataset from the official website, and the official train/test split from here. Put them under the same folder "$NYUFOLDER". To evaluate, please run, ```bash python evaluate.py --methods=f --suffix=warpingrefine --dataset=nyu --dataFolder="$NYU_FOLDER" ```

Note that the numbers are off with the provided model. We retrained the model after cleaning up the code, which is different from the model we tested for the publication.

Contact

If you have any questions, please contact the primary author Chen Liu <[email protected]>, or Kihwan Kim <[email protected]>.

Acknowledgement

Our implementation uses the nms/roialign from the Mask R-CNN implementation from pytorch-mask-rcnn, which is licensed under MIT License

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.