Semantic segmentation of 2D floor plans with a pixel-wise classifier

Patent metadata

Field	Value
Publication number	US-12051261-B2
Application number	US-201816235930-A
Country	US
Kind code	B2
Filing date	Dec 28, 2018
Priority date	Dec 28, 2017
Publication date	Jul 30, 2024
Grant date	Jul 30, 2024

Abstract

The disclosure notably relates to a computer-implemented method for determining a function configured to determine a semantic segmentation of a 2D floor plan representing a layout of a building. The method comprises providing a dataset comprising 2D floor plans each associated to a respective semantic segmentation. The method also comprises learning the function based on the dataset. Such a method provides an improved solution for processing a 2D floor plan.

First claim

The invention claimed is:

1. A computer-implemented method for determining a function configured to determine a semantic segmentation of a 2D floor plan representing a layout of a building, the function having a neural network presenting a convolutional encoder-decoder architecture, the neural network comprising a pixel-wise classifier with respect to a set of classes comprising a wall class, a door class and a window class, the method comprising:
obtaining a dataset comprising 2D floor plans each associated to a respective semantic segmentation; and
learning the function based on the dataset,
wherein the pixel-wise classifier is a single layer of the neural network consisting of a softmax classifier, the neural network includes weights, the neural network further comprising an upsampling layer being arranged before the softmax classifier and a max-pooling layer being arranged after the softmax classifier, and the learning includes, with a single optimization algorithm, updating the weights according to the dataset and to a loss function, wherein the optimization algorithm is a stochastic gradient descent, the weights of the neural network being definitely set after said stochastic gradient descent, including the weights of the softmax classifier and the max-pooling layer, the weights of the max-pooling layer being optimized by the learning,
wherein the softmax classifier outputs, for each input 2D floor plan, respective data for inference of a semantic segmentation mask of the input 2D floor plan, the semantic segmentation mask being a pixel-wise classification of the 2D floor plan with respect to the set of classes, the loss function penalizing, for each 2D floor plan of the dataset, inference of a semantic segmentation mask erroneous relative to the respective semantic segmentation associated to the 2D floor plan in the dataset,
wherein the softmax classifier outputs, for each pixel of an input 2D floor plan, respective data for inference of a class of the set of classes, the loss function penalizing, for each pixel of each 2D floor plan of the dataset, inference of a respective class different from a class provided for said pixel by the respective semantic segmentation associated to the 2D floor plan in the dataset, a determined class being the class with the highest probability among the distribution of probabilities comprised in the respective data outputted by the softmax classifier, the function thereby presenting a mean accuracy higher than 0.85 and/or a mean intersection-over-union higher than 0.75, such that ambiguities between the door class and the window class are reduced,
wherein the respective data outputted by the softmax classifier comprises a distribution of probabilities over the set of classes, the softmax classifier thereby outputting a respective distribution of probabilities over the set of classes for each pixel, the loss function comprising a sum of loss terms each relative to a respective pixel, each loss term being of the type:

-

∑

i
=
1

C

⁢

⁢

y
true
i

⁢

log
⁡

(

y
pred
i

)

where:
C is the number of classes of the set of classes comprising the wall class, the door class and the window class,
i designates a class of the set of classes,
y true i is a binary indicator if class i is the class provided for the respective pixel by the respective semantic segmentation associated to the 2D floor plan in the dataset, and
y pred i is a probability outputted by the softmax classifier for class i,

wherein the loss function is multinomial.

2. The method of claim 1 , wherein the neural network comprises weights, and the learning comprises, with an optimization algorithm, updating the weights according to the dataset and to a loss function.

3. The method of claim 2 , wherein the optimization algorithm is a stochastic gradient descent.

4. The method of claim 3 , wherein the loss function is a cross-entropy loss function.

5. The method of claim 1 , wherein providing the dataset comprises:
obtaining a database of 2D floor plans each associated to a respective 3D model; and
determining for each 2D floor plan the respective semantic segmentation from the respective 3D model.

6. The method of claim 1 , the set of classes consisting of the wall class, the door class and the window class.

7. A computer-implemented method comprising:
determining a semantic segmentation of a 2D floor plan representing a layout of a building, by:
obtaining the 2D floor plan, and
applying a function to the 2D floor plan, the function being learnable according to a computer-implemented process for determining a function configured to determine a semantic segmentation of a 2D floor plan representing a layout of a building, the function having a neural network presenting a convolutional encoder-decoder architecture, the neural network comprising a pixel-wise classifier with respect to a set of classes comprising a wall class, a door class and a window class, the process including:
obtaining a dataset comprising 2D floor plans each associated to a respective semantic segmentation, and
learning the function based on the dataset,

wherein the pixel-wise classifier is a single layer of the neural network consisting of a softmax classifier, the neural network includes weights, the neural network further comprising an upsampling layer being arranged before the softmax classifier and a max-pooling layer being arranged after the softmax classifier, and the learning includes, with a single optimization algorithm, updating the weights according to the dataset and to a loss function, wherein the optimization algorithm is a stochastic gradient descent, the weights of the neural network being definitely set after said stochastic gradient descent, including the weights of the softmax classifier and the max-pooling layer, the weights of the max-pooling layer being optimized by the learning,
wherein the softmax classifier outputs, for each input 2D floor plan, respective data for inference of a semantic segmentation mask of the input 2D floor plan, the semantic segmentation mask being a pixel-wise classification of the 2D floor plan with respect to the set of classes, the loss function penalizing, for each 2D floor plan of the dataset, inference of a semantic segmentation mask erroneous relative to the respective semantic segmentation associated to the 2D floor plan in the dataset,
wherein the softmax classifier outputs, for each pixel of an input 2D floor plan, respective data for inference of a class of the set of classes, the loss function penalizing, for each pixel of each 2D floor plan of the dataset, inference of a respective class different from a class provided for said pixel by the respective semantic segmentation associated to the 2D floor plan in the dataset, a determined class being the class with the highest probability among the distribution of probabilities comprised in the respective data outputted by the

Assignees

• Dassault Systemes

Inventors

• Rejeb Sfar Asma
• Dupont De Dinechin Louis
• Boulkenafed Malika

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title