Intelligent Home 3D: Automatic 3D-House Design From Linguistic Descriptions Only

Chen, Qi; Wu, Qi; Tang, Rui; Wang, Yuhan; Wang, Shuai; Tan, Mingkui

doi:10.1109/cvpr42600.2020.01264

Cited by 31 publications

(24 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Text2Scene [24] solves a similar task in 2 dimensions, by iteratively placing objects into an 2D image and then ensuring consistency. Intelligent Home 3D [9] tackles the related task of generating the full room layout of a house from text, and proposes a new dataset for this task. Our method differs from these in that it can generate high quality complex scenes with a large number of objects, without requiring user inputs for refinement.…”

Section: Related Workmentioning

confidence: 99%

SceneFormer: Indoor Scene Generation with Transformers

Wang

Yeshwanth

Niesner

2021

2021 International Conference on 3D Vision (3DV)

View full text Add to dashboard Cite

Figure 1: Shape-conditioned scenes generated by our bedroom and living room models. In each pair, the image on the left is a visualization of the input to the model that includes the shape of the floor of the room, and locations of the doors and windows. The image on the right is the generated scene, with objects added. Our models learn the desired locations of salient objects in each task -the bed and sofa, and then add other objects that augment the scene in a realistic manner. Importantly, we can generate objects on the walls and ceilings for overall realism, as compared to only objects on the floor in existing works.

show abstract

Section: Related Workmentioning

confidence: 99%

SceneFormer: Indoor Scene Generation with Transformers

Wang

Yeshwanth

Niesner

2021

2021 International Conference on 3D Vision (3DV)

View full text Add to dashboard Cite

show abstract

“…Testing framework: To avoid a network from simply copying and pasting layouts, we use the k-fold cross validation from House-GAN [24], while dividing the samples into four groups based on the number of rooms: (5,6,7,8): When generating layouts with 8 rooms, we use samples from the other three groups for training. At test time, we randomly pick a GT layout from the test set, use its bubble-diagram to initialize a relational graph, and generate a layout sample.…”

Section: Implementation Detailsmentioning

confidence: 99%

“…Early methods formulated it as iterative optimization [22,21]. The surge of deep neural networks has made a breakthrough, where state-of-the-art algorithms [24,7] utilize Generative Adversarial Networks (GANs) [10]. GAN is a one-shot generation process, converting a noise vector into a sample.…”

Section: Introductionmentioning

confidence: 99%

House-GAN++: Generative Adversarial Layout Refinement Networks

Nauata,

Hosseini,

Chang

et al. 2021

Preprint

View full text Add to dashboard Cite

Figure 1. The paper makes a breakthrough in the task of automated house layout generation. The right shows the mix of a ground-truth design made by an architect and our generated samples, based on the input bubble-diagram. Can you tell which one is the ground-truth? See the end of the caption for the answer. The paper proposes a novel generative adversarial layout refinement network, whose generator is trained to repeatedly apply and refine the design towards perfection. (The second sample from the right is the ground-truth.

show abstract

“…The differences can be noticed if we take into account not only the method used but also the type of input that the method accepts. These input constraints can be Graph-based, which take the form of bubble diagrams as input (Hu et al, 2020;Nauata et al, 2021;Wu et al, 2019), Language-based, which takes linguistic descriptions as input to the generative model (Chen et al, 2020;Galanos, 2021), and last but not least Pixel-based approaches, which use the pixel color as constrains to the generative model, whereas information like shape, orientation or area could be further determined (Chaillou, 2020;Peters, 2018;Rahbar et al, 2019).…”

Section: Introductionmentioning

confidence: 99%

Big Data vs Smart Data on the Generation of Floor Plans with Deep Learning

Rodrigues¹,

Imagawa²,

Koga³

et al. 2021

Blucher Design Proceedings

View full text Add to dashboard Cite

Due to the progressive growth of data dimensionality, addressing how much data and time is required to train deep learning models has become an important research topic. Thus, in this paper, we present a benchmark for generating floor plans with Conditional Generative Adversarial Networks in which we compare 10 trained models on a dataset of 80.000 samples, the models use different data dimensions and hyper-parameters on the training phase, beyond this objective, we also tested the capability of Convolutional Neural Networks (CNN) to reduce the dataset noise. The models' assessment was made on more than 6 million with the Frétche Inception Distance (FID). The results show that such models can rapidly achieve similar or even better FID results if trained with 800 images of 512x512 pixels, in comparison to high dimensional datasets of 256x256 pixels, however, using CNNs to enhance data consistency reproduced optimal results using around 27.000 images.

show abstract

Intelligent Home 3D: Automatic 3D-House Design From Linguistic Descriptions Only

Cited by 31 publications

References 29 publications

SceneFormer: Indoor Scene Generation with Transformers

SceneFormer: Indoor Scene Generation with Transformers

House-GAN++: Generative Adversarial Layout Refinement Networks

Big Data vs Smart Data on the Generation of Floor Plans with Deep Learning

Contact Info

Product

Resources

About