Orientation-boosted Voxel Nets for 3D Object Recognition

By Nima Sedaghat, Mohammadreza Zolfaghari, Thomas Brox. 2016

233 citations recorded by the source corpus.

Recent work has shown good recognition results in 3D object recognition using 3D convolutional networks. In this paper, we show that the object orientation plays an important role in 3D recognition. More specifically, we argue that objects induce different features in the network under rotation. Thus, we approach the category-level classification task as a multi-task problem, in which the network is trained to predict the pose of the object in addition to the class label as a parallel task. We show that this yields significant improvements in the classification results. We test our suggested architecture on several datasets representing various 3D data sources: LiDAR data, CAD models, and RGB-D images. We report state-of-the-art results on classification as well as significant improvements in precision and speed over the baseline on 3D detection.

Subfields: Computational Mechanics, Computer Vision and Pattern Recognition, 3D Vision & Pose

How researchers see this paper

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

In Computational Mechanics

In Computer Vision and Pattern Recognition

In 3D Vision & Pose

Citations in the corpus

OctNet: Learning Deep 3D Representations at High Resolutions (2016, background): “[1] showed that introducing an auxiliary orientation loss increases classification performance over the original VoxNet.”
Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs (2017, methodology): “We have introduced edge-conditioned convolution (ECC), an operation on graph signal performed in the spatial domain where filter weights are conditioned on edge labels and dynamically generated for each specific input sample.”
Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling (2016, neutral): “Results We compare with the state-of-the-art methods [Wu et al., 2015, Girdhar et al., 2016, Sharma et al., 2016, Sedaghat et al., 2016] and show per-class accuracy in Table 1.”
Generative and Discriminative Voxel Modeling with Convolutional Neural Networks (2016, methodology): “Model ModelNet40 ModelNet10 Voxnet [5] 83.00% 92.00% MVCNN [2] 90.10% - Pairwise [3] 90.70% 92.80% FusionNets [4] 90.80% 93.11% ORION [6] - 93.80% MVCNN-Multires [22] 91.40 % - Voxception 90.56% 93.28% VRN (One-View) 88.98% - VRN 91.33% 93.61% VRN Ensemble 95.54% 97.14% with consistent structure, in”
Polar Transformer Networks (2017, positive): “gmenting the training data with transformed versions of the inputs (Wu et al., 2015), or pooling over transformed versions during training and/or test (Maturana & Scherer, 2015; Qi et al., 2016). Sedaghat et al. (2016) show that a multi-task approach, i.e. prediction of both the orientation and ”
CubeNet: Equivariance to 3D Rotation and Translation (2018, methodology): “Interestingly, our single-view result we obtain is very similar to ORION [43], which introduces an orientation estimation task along with the classification.”
SPLATNet: Sparse Lattice Networks for Point Cloud Processing (2018, background): “Early voxel-based architectures executed convolution in regular, fixed voxel grids, and were limited to low shape resolutions due to high memory and computation costs [43, 28, 32, 6, 15, 37].”
3D Object Classification via Spherical Projections (2017, methodology): “ORION [2] adds an orientation estimation module to the original VoxNet [13] as another task and trains both tasks simultaneously, which boosts the performance of VoxNet.”
Triplet-Center Loss for Multi-view 3D Object Retrieval (2018, methodology): “3D model-based methods directly learn shape features from 3D data formats, such as polygon meshes or surfaces [5, 6, 42, 43], voxel grid [21, 40, 18, 25, 30], and point clouds [24, 26].”
Generating 3D Adversarial Point Clouds (2018, methodology): “We use the aligned benchmark ModelNet40 [27, 23] dataset for our experiments.”
View Inter-Prediction GAN: Unsupervised Representation Learning for 3D Shapes by Learning Global Shape Memories to Support Local View Predictions (2018, methodology): “The compared methods include MVCNN (Su et al. 2015), ORION (Sedaghat et al. 2017), 3DDescriptorNet (Xie et al. 2018), Pair-wise (Johns, Leutenegger, and Davison 2016), GIFT (Bai et al. 2017), PANORAMA (Sﬁkas, Theoharis, and Pratikakis 2017), VRN (Brock et al. 2016), Rotation-Net (Kanezaki,…”
General-Purpose Deep Point Cloud Feature Extractor (2018, methodology): “[39, 41] have created a method of automatically aligning objects per-class that we use to reduce the variance between ModelNet40 samples.”
Noise-Resistant Deep Learning for Object Classification in Three-Dimensional Point Clouds Using a Point Pair Descriptor (2018, background): “clouds into 3D voxel grids ([11] and [12], to name a few).”