Seeing 3D Chairs: Exemplar Part-Based 2D-3D Alignment Using a Large Dataset of CAD Models

By Mathieu Aubry, Daniel Maturana, Alexei A. Efros, Bryan C. Russell, Josef Sivic. 2014

553 citations recorded by the source corpus.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition, Geology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

In Computer Vision and Pattern Recognition

In Geology

Citations in the corpus

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets (2016, methodology): “Next we evaluate InfoGAN on two datasets of 3D images: faces [28] and chairs [29], on which DC-IGN was shown to learn highly interpretable graphics codes.”
Learning to generate chairs with convolutional neural networks (2014, methodology): “Given the set of 3D chair models of Aubry et al. [1], we aim to train a neural network capable of generating 2D projections of the models given the chair type, viewpoint, and, optionally, other parameters such as color, brightness, saturation, zoom, etc.”
Deep Convolutional Inverse Graphics Network (2015, methodology): “tiated and represented by z[2;200]. The DC-IGN network succeeded in achieving a mean-squared error (MSE) of reconstruction of 2:7722 10 4 on the test set. Each image has grayscale values in the range [0;1] and is 150 150 pixels. In Figure 7 we have included examples of the networks ability to re-ren”
Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (2017, methodology): “In this sense, similar to what has been shown in previous Car to Car experiment, the two models suffered from mode collapse.”
Unsupervised Adversarial Invariance (2018, background): “hance. 5.1 Invariance to inherent nuisance factors We provide results of our framework at the task of learning invariance to inherent nuisance factors on two datasets – Extended Yale-B [7] and Chairs [2]. Extended Yale-B. This dataset contains face-images of 38 subjects under various lighting condit”
Visual interpretability for deep learning: a survey (2018, methodology): “The InfoGAN has been trained using the MNIST dataset [LeCun et al., 1998b], the CelebA dataset [Liu et al., 2015], the SVHN dataset [Netzer et al., 2011], the 3D face dataset [Paysan et al., 2009], and the 3D chair dataset [Aubry et al., 2014].”
Disentangling by Factorising (2018, background): “4) 3D Chairs (Aubry et al., 2014): 86,366 RGB 64× 64× 3 images of chair CAD models.”
Isolating Sources of Disentanglement in VAEs (2018, background): “We show qualitatively that β-TCVAE discovers more disentangled factors than β-VAE on datasets of chairs [41] and real faces [42].”
The Challenges of Continuous Self-Supervised Learning (2022, background): “These techniques are made possible by the fact that training data is constructed by collecting samples from images in significantly different datasets - for example, [82] uses Celeb-A[40] faces followed by 3D-Chair[5] images).”
Learning latent representations across multiple data domains using Lifelong VAEGAN (2020, background): “consider a learning rate of 0.001. The generation and reconstruction results are presented in Figs. 2a-c, for the lifelong learning of CelebA [39] to CACD [9], and in Figs. 2d-f for CelebA to 3DChair [6]. From these results, the inference model works well for both discrete and continuous latent vari”
Pix3D: Dataset and Methods for Single-Image 3D Shape Modeling (2018, background): “Another related research direction is retrieving similar 3D shapes given a single image, instead of reconstructing the object’s actual geometry [1, 15, 19, 48].”
Learning a Predictable and Generative Vector Representation for Objects (2016, methodology): “Many approaches achieve this via alignment [21,22,23]; the most similar work is Li et al. [24], which learns a joint, but non-generative, embedding of images and 3D shapes.”
Lifting Object Detection Datasets into 3D (2015, neutral): “To account for differences between the 3D exemplars and the object depicted, Su and Guibas [26] choose a few exemplars that appear to be similar to the depicted shape and combine them to produce a single depth map.”
FlowNet: Learning Optical Flow with Convolutional Networks (2015, methodology): “for quantity, we generate a synthetic Flying Chairs dataset which consists of random background images from Flickr on which we overlay segmented images of chairs from [1].”
Variational Inference of Disentangled Latent Concepts from Unlabeled Observations (2017, background): “To be consistent with the evaluations in (Higgins et al., 2017), we use the same CNN network architectures (for our encoder and decoder), and same latent dimensions as used in (Higgins et al., 2017) for CelebA, 3D Chairs, 2D Shapes datasets.”
Understanding disentangling in β-VAE (2018, background): “ata samples used to generate them (top row). This contrasts with the results previously obtained with the ﬁxed -modulated KL objective in [15]. We also trained the same model on the 3D Chairs dataset [3], with latent traversals shown in Figure 4b. We can see that reconstructions are of high quality,”
Weakly-supervised Disentangling with Recurrent Transformations for 3D View Synthesis (2015, methodology): “We tune the K value in [1,3,5,7], namely KNN1, KNN3, KNN5 and KNN7 to achieve its best performance.”
Playing for Data: Ground Truth from Computer Games (2016, neutral): “3D object models are also increasingly used for training representations for object detection and object pose estimation [4, 26, 30, 34, 35, 45].”
Single Image 3D Interpreter Network (2016, methodology): “ength, and seat width, and extrinsic properties such as its pose. In this paper, we recover these object properties from a single image by estimating 3D structure. Instead of a 3D mesh or a depth map [9,2,32,18,40,50,16], we represent an object via a 3D skeleton [47], which consists of keypoints and”
Render for CNN: Viewpoint Estimation in Images Using CNNs Trained with Rendered 3D Model Views (2015, methodology): “Leveraging 3D repositories, [20, 1] use 250 and 1.”