test自己的数据时，需要多大内存？

FloorNet: A Unified Framework for Floorplan Reconstruction from 3D Scans

By Chen Liu^*, Jiaye Wu^*, and Yasutaka Furukawa (^* indicates equal contribution)

Introduction

This paper proposes FloorNet, a novel neural network, to turn RGBD videos of indoor spaces into vector-graphics floorplans. FloorNet consists of three branches, PointNet branch, Floorplan branch, and Image branch. For more details, please refer to our ECCV 2018 paper or visit our project website. This is a follow-up work of our floorplan transformation project which you can find here.

Updates

[12/22/2018] We now provide a free IP solver (not relying on Gurobi) at IP.py. The functionality of IP.py should be similar to QP.py which uses Gurobi to solve the IP problem. You might want to consider the free solver if you don't have a Gurobi license.

Dependencies

Python 2.7, TensorFlow (>= 1.3), numpy, opencv 3, CUDA (>= 8.0), Gurobi (free only for academic usages).

Data

Dataset used in the paper

We collect 155 scans of residential units and annotated corresponding floorplan information. Among 155 scans, 135 are used for training and 20 are for testing. We convert data to tfrecords files which can be downloaded here (or here if you cannot access the previous one). Please put the downloaded files under folder data/.

Here are the links to the raw point clouds, annotations, and their associations. Please refer to RecordWriterTango.py to see how to convert the raw data and annotations to tfrecords files.

Using custom data

To generate training/testing data from other data source, the data should be converted to tfrecords as what we did in RecordWriterTango.py (an example of our raw data before processed by RecordWriterTango.py is provided here). Please refer to this guide for how to generate and read tfrecords.

Basically, every data sample(tf.train.Example) should at least contain the following components:

Inputs:
- a point cloud (randomly sampled 50,000 points)
- a mapping from point cloud's 3D space to 2D space of the 256x256 top-view density image.
  - It contains 50,000 indices, one for each point.
  - For point (x, y, z), index = round((y - min(Y) + padding) / (maxRange + 2 * padding) * 256) * 256 + round((x - min(X) + padding) / (maxRange + 2 * padding) * 256).
    - maxRange = max(max(X) - min(X), max(Y) - min(Y))
    - padding could be any small value, say 0.05 maxRange
- optional: image features of the RGB video stream, if the image branch is enabled
Labels:
- Corners and their corresponding types
- Total number of corners
- A ground-truth icon segmentation map
- A ground-truth room segmentation map

Again, please refer to RecordWriterTango.py for exact details.

NEW: We added a template file, RecordWriterCustom.py for using custom data.

Annotator

For reference, a similar (but not the same) annotator written in Python is here. You need to make some changes to annotate your own data.

Training

To train the network from scratch, please run:

python train.py --restore=0

Evaluation

To evaluate the performance of our trained model, please run:

python train.py --task=evaluate --separateIconLoss

Generate 3D models

We can popup the reconstructed floorplan to generate 3D models. Please refer to our previous project, FloorplanTransformation, for more details.

Contact

If you have any questions, please contact me at [email protected].

art-programmer / floornet Goto Github PK

floornet's Introduction

FloorNet: A Unified Framework for Floorplan Reconstruction from 3D Scans

Introduction

Updates

Dependencies

Data

Dataset used in the paper

Using custom data

Annotator

Training

Evaluation

Generate 3D models

Contact

floornet's People

Contributors

Stargazers

Watchers

Forkers

floornet's Issues

Recommend Projects

Recommend Topics

Recommend Org