Zhizhong Han

Junsheng Zhou, Baorui Ma, Shujuan Li et al. · tsinghua

Latest methods represent shapes with open surfaces using unsigned distance functions (UDFs). They train neural networks to learn UDFs and reconstruct surfaces with the gradients around the zero level set of the UDF. However, the differential networks struggle from learning the zero level set where the UDF is not differentiable, which leads to large errors on unsigned distances and gradients around the zero level set, resulting in highly fragmented and discontinuous surfaces. To resolve this problem, we propose to learn a more continuous zero level set in UDFs with level set projections. Our insight is to guide the learning of zero level set using the rest non-zero level sets via a projection procedure. Our idea is inspired from the observations that the non-zero level sets are much smoother and more continuous than the zero level set. We pull the non-zero level sets onto the zero level set with gradient constraints which align gradients over different level sets and correct unsigned distance errors on the zero level set, leading to a smoother and more continuous unsigned distance field. We conduct comprehensive experiments in surface reconstruction for point clouds, real scans or depth maps, and further explore the performance in unsupervised point cloud upsampling and unsupervised point normal estimation with the learned UDF, which demonstrate our non-trivial improvements over the state-of-the-art methods. Code is available at https://github.com/junshengzhou/LevelSetUDF .

Chao Chen, Yu-Shen Liu, Zhizhong Han · tsinghua

Learning implicit representations has been a widely used solution for surface reconstruction from 3D point clouds. The latest methods infer a distance or occupancy field by overfitting a neural network on a single point cloud. However, these methods suffer from a slow inference due to the slow convergence of neural networks and the extensive calculation of distances to surface points, which limits them to small scale points. To resolve the scalability issue in surface reconstruction, we propose GridPull to improve the efficiency of learning implicit representations from large scale point clouds. Our novelty lies in the fast inference of a discrete distance field defined on grids without using any neural components. To remedy the lack of continuousness brought by neural networks, we introduce a loss function to encourage continuous distances and consistent gradients in the field during pulling queries onto the surface in grids near to the surface. We use uniform grids for a fast grid search to localize sampled queries, and organize surface points in a tree structure to speed up the calculation of distances to the surface. We do not rely on learning priors or normal supervision during optimization, and achieve superiority over the latest methods in terms of complexity and accuracy. We evaluate our method on shape and scene benchmarks, and report numerical and visual comparisons with the latest methods to justify our effectiveness and superiority. The code is available at https://github.com/chenchao15/GridPull.

Peng Xiang, Xin Wen, Yu-Shen Liu et al. · tsinghua

Learning per-point semantic features from the hierarchical feature pyramid is essential for point cloud semantic segmentation. However, most previous methods suffered from ambiguous region features or failed to refine per-point features effectively, which leads to information loss and ambiguous semantic identification. To resolve this, we propose Retro-FPN to model the per-point feature prediction as an explicit and retrospective refining process, which goes through all the pyramid layers to extract semantic features explicitly for each point. Its key novelty is a retro-transformer for summarizing semantic contexts from the previous layer and accordingly refining the features in the current stage. In this way, the categorization of each point is conditioned on its local semantic pattern. Specifically, the retro-transformer consists of a local cross-attention block and a semantic gate unit. The cross-attention serves to summarize the semantic pattern retrospectively from the previous layer. And the gate unit carefully incorporates the summarized contexts and refines the current semantic features. Retro-FPN is a pluggable neural network that applies to hierarchical decoders. By integrating Retro-FPN with three representative backbones, including both point-based and voxel-based methods, we show that Retro-FPN can significantly improve performance over state-of-the-art backbones. Comprehensive experiments on widely used benchmarks can justify the effectiveness of our design. The source is available at https://github.com/AllenXiangX/Retro-FPN

Sijia Jiang, Jing Hua, Zhizhong Han

In recent years, huge progress has been made on learning neural implicit representations from multi-view images for 3D reconstruction. As an additional input complementing coordinates, using sinusoidal functions as positional encodings plays a key role in revealing high frequency details with coordinate-based neural networks. However, high frequency positional encodings make the optimization unstable, which results in noisy reconstructions and artifacts in empty space. To resolve this issue in a general sense, we introduce to learn neural implicit representations with quantized coordinates, which reduces the uncertainty and ambiguity in the field during optimization. Instead of continuous coordinates, we discretize continuous coordinates into discrete coordinates using nearest interpolation among quantized coordinates which are obtained by discretizing the field in an extremely high resolution. We use discrete coordinates and their positional encodings to learn implicit functions through volume rendering. This significantly reduces the variations in the sample space, and triggers more multi-view consistency constraints on intersections of rays from different views, which enables to infer implicit function in a more effective way. Our quantized coordinates do not bring any computational burden, and can seamlessly work upon the latest methods. Our evaluations under the widely used benchmarks show our superiority over the state-of-the-art. Our code is available at https://github.com/MachinePerceptionLab/CQ-NIR.

Junsheng Zhou, Baorui Ma, Yu-Shen Liu et al.

Learning signed distance functions (SDFs) from point clouds is an important task in 3D computer vision. However, without ground truth signed distances, point normals or clean point clouds, current methods still struggle from learning SDFs from noisy point clouds. To overcome this challenge, we propose to learn SDFs via a noise to noise mapping, which does not require any clean point cloud or ground truth supervision. Our novelty lies in the noise to noise mapping which can infer a highly accurate SDF of a single object or scene from its multiple or even single noisy observations. We achieve this by a novel loss which enables statistical reasoning on point clouds and maintains geometric consistency although point clouds are irregular, unordered and have no point correspondence among noisy observations. To accelerate training, we use multi-resolution hash encodings implemented in CUDA in our framework, which reduces our training time by a factor of ten, achieving convergence within one minute. We further introduce a novel schema to improve multi-view reconstruction by estimating SDFs as a prior. Our evaluations under widely-used benchmarks demonstrate our superiority over the state-of-the-art methods in surface reconstruction from point clouds or multi-view images, point cloud denoising and upsampling.

Qing Li, Huifang Feng, Kanle Shi et al.

Estimating the normal of a point requires constructing a local patch to provide center-surrounding context, but determining the appropriate neighborhood size is difficult when dealing with different data or geometries. Existing methods commonly employ various parameter-heavy strategies to extract a full feature description from the input patch. However, they still have difficulties in accurately and efficiently predicting normals for various point clouds. In this work, we present a new idea of feature extraction for robust normal estimation of point clouds. We use the fusion of multi-scale features from different neighborhood sizes to address the issue of selecting reasonable patch sizes for various data or geometries. We seek to model a patch feature fitting (PFF) based on multi-scale features to approximate the optimal geometric description for normal estimation and implement the approximation process via multi-scale feature aggregation and cross-scale feature compensation. The feature aggregation module progressively aggregates the patch features of different scales to the center of the patch and shrinks the patch size by removing points far from the center. It not only enables the network to precisely capture the structure characteristic in a wide range, but also describes highly detailed geometries. The feature compensation module ensures the reusability of features from earlier layers of large scales and reveals associated information in different patch sizes. Our approximation strategy based on aggregating the features of multiple scales enables the model to achieve scale adaptation of varying local patches and deliver the optimal feature description. Extensive experiments demonstrate that our method achieves state-of-the-art performance on both synthetic and real-world datasets with fewer network parameters and running time.

Chao Chen, Yu-Shen Liu, Zhizhong Han

It is important to estimate an accurate signed distance function (SDF) from a point cloud in many computer vision applications. The latest methods learn neural SDFs using either a data-driven based or an overfitting-based strategy. However, these two kinds of methods are with either poor generalization or slow convergence, which limits their capability under challenging scenarios like highly noisy point clouds. To resolve this issue, we propose a method to promote pros of both data-driven based and overfitting-based methods for better generalization, faster inference, and higher accuracy in learning neural SDFs. We introduce a novel statistical reasoning algorithm in local regions which is able to finetune data-driven based priors without signed distance supervision, clean point cloud, or point normals. This helps our method start with a good initialization, and converge to a minimum in a much faster way. Our numerical and visual comparisons with the state-of-the-art methods show our superiority over these methods in surface reconstruction and point cloud denoising on widely used shape and scene benchmarks. The code is available at https://github.com/chenchao15/LocalN2NM.

Baorui Ma, Junsheng Zhou, Yu-Shen Liu et al.

Neural signed distance functions (SDFs) have shown remarkable capability in representing geometry with details. However, without signed distance supervision, it is still a challenge to infer SDFs from point clouds or multi-view images using neural networks. In this paper, we claim that gradient consistency in the field, indicated by the parallelism of level sets, is the key factor affecting the inference accuracy. Hence, we propose a level set alignment loss to evaluate the parallelism of level sets, which can be minimized to achieve better gradient consistency. Our novelty lies in that we can align all level sets to the zero level set by constraining gradients at queries and their projections on the zero level set in an adaptive way. Our insight is to propagate the zero level set to everywhere in the field through consistent gradients to eliminate uncertainty in the field that is caused by the discreteness of 3D point clouds or the lack of observations from multi-view images. Our proposed loss is a general term which can be used upon different methods to infer SDFs from 3D point clouds and multi-view images. Our numerical and visual comparisons demonstrate that our loss can significantly improve the accuracy of SDFs inferred from point clouds or multi-view images under various benchmarks. Code and data are available at https://github.com/mabaorui/TowardsBetterGradient .

Peng Xiang, Xin Wen, Yu-Shen Liu et al.

Point cloud completion aims to predict a complete shape in high accuracy from its partial observation. However, previous methods usually suffered from discrete nature of point cloud and unstructured prediction of points in local regions, which makes it hard to reveal fine local geometric details on the complete shape. To resolve this issue, we propose SnowflakeNet with Snowflake Point Deconvolution (SPD) to generate the complete point clouds. The SnowflakeNet models the generation of complete point clouds as the snowflake-like growth of points in 3D space, where the child points are progressively generated by splitting their parent points after each SPD. Our insight of revealing detailed geometry is to introduce skip-transformer in SPD to learn point splitting patterns which can fit local regions the best. Skip-transformer leverages attention mechanism to summarize the splitting patterns used in the previous SPD layer to produce the splitting in the current SPD layer. The locally compact and structured point cloud generated by SPD is able to precisely capture the structure characteristic of 3D shape in local patches, which enables the network to predict highly detailed geometries, such as smooth regions, sharp edges and corners. Our experimental results outperform the state-of-the-art point cloud completion methods under widely used benchmarks. Code will be available at https://github.com/AllenXiangX/SnowflakeNet.

Han Liu, Zhizhong Han, Yu-Shen Liu et al.

Low-rank metric learning aims to learn better discrimination of data subject to low-rank constraints. It keeps the intrinsic low-rank structure of datasets and reduces the time cost and memory usage in metric learning. However, it is still a challenge for current methods to handle datasets with both high dimensions and large numbers of samples. To address this issue, we present a novel fast low-rank metric learning (FLRML) method.FLRML casts the low-rank metric learning problem into an unconstrained optimization on the Stiefel manifold, which can be efficiently solved by searching along the descent curves of the manifold.FLRML significantly reduces the complexity and memory usage in optimization, which makes the method scalable to both high dimensions and large numbers of samples.Furthermore, we introduce a mini-batch version of FLRML to make the method scalable to larger datasets which are hard to be loaded and decomposed in limited memory. The outperforming experimental results show that our method is with high accuracy and much faster than the state-of-the-art methods under several benchmarks with large numbers of high-dimensional data. Code has been made available at https://github.com/highan911/FLRML

Pengchong Hu, Zhizhong Han

Learning neural implicit representations has achieved remarkable performance in 3D reconstruction from multi-view images. Current methods use volume rendering to render implicit representations into either RGB or depth images that are supervised by multi-view ground truth. However, rendering a view each time suffers from incomplete depth at holes and unawareness of occluded structures from the depth supervision, which severely affects the accuracy of geometry inference via volume rendering. To resolve this issue, we propose to learn neural implicit representations from multi-view RGBD images through volume rendering with an attentive depth fusion prior. Our prior allows neural networks to perceive coarse 3D structures from the Truncated Signed Distance Function (TSDF) fused from all depth images available for rendering. The TSDF enables accessing the missing depth at holes on one depth image and the occluded parts that are invisible from the current view. By introducing a novel attention mechanism, we allow neural networks to directly use the depth fusion prior with the inferred occupancy as the learned implicit function. Our attention mechanism works with either a one-time fused TSDF that represents a whole scene or an incrementally fused TSDF that represents a partial scene in the context of Simultaneous Localization and Mapping (SLAM). Our evaluations on widely used benchmarks including synthetic and real-world scans show our superiority over the latest neural implicit methods. Project page: https://machineperceptionlab.github.io/Attentive_DF_Prior/

Sijia Jiang, Tong Wu, Jing Hua et al.

It is vital to recover 3D geometry from multi-view RGB images in many 3D computer vision tasks. The latest methods infer the geometry represented as a signed distance field by minimizing the rendering error on the field through volume rendering. However, it is still challenging to explicitly impose constraints on surfaces for inferring more geometry details due to the limited ability of sensing surfaces in volume rendering. To resolve this problem, we introduce a method to infer signed distance functions (SDFs) with a better sense of surfaces through volume rendering. Using the gradients and signed distances, we establish a small surface patch centered at the estimated intersection along a ray by pulling points randomly sampled nearby. Hence, we are able to explicitly impose surface constraints on the sensed surface patch, such as multi-view photo consistency and supervision from depth or normal priors, through volume rendering. We evaluate our method by numerical and visual comparisons on scene benchmarks. Our superiority over the latest methods justifies our effectiveness.

Sijia Jiang, Jing Hua, Zhizhong Han

Neural implicit representations have shown remarkable abilities in jointly modeling geometry, color, and camera poses in simultaneous localization and mapping (SLAM). Current methods use coordinates, positional encodings, or other geometry features as input to query neural implicit functions for signed distances and color which produce rendering errors to drive the optimization in overfitting image observations. However, due to the run time efficiency requirement in SLAM systems, we are merely allowed to conduct optimization on each frame in few iterations, which is far from enough for neural networks to overfit these queries. The underfitting usually results in severe drifts in camera tracking and artifacts in reconstruction. To resolve this issue, we propose query quantized neural SLAM which uses quantized queries to reduce variations of input for much easier and faster overfitting a frame. To this end, we quantize a query into a discrete representation with a set of codes, and only allow neural networks to observe a finite number of variations. This allows neural networks to become increasingly familiar with these codes after overfitting more and more previous frames. Moreover, we also introduce novel initialization, losses, and argumentation to stabilize the optimization with significant uncertainty in the early optimization stage, constrain the optimization space, and estimate camera poses more accurately. We justify the effectiveness of each design and report visual and numerical comparisons on widely used benchmarks to show our superiority over the latest methods in both reconstruction and camera tracking.

Qing Li, Huifang Feng, Kanle Shi et al.

We propose a novel method called SHS-Net for oriented normal estimation of point clouds by learning signed hyper surfaces, which can accurately predict normals with global consistent orientation from various point clouds. Almost all existing methods estimate oriented normals through a two-stage pipeline, i.e., unoriented normal estimation and normal orientation, and each step is implemented by a separate algorithm. However, previous methods are sensitive to parameter settings, resulting in poor results from point clouds with noise, density variations and complex geometries. In this work, we introduce signed hyper surfaces (SHS), which are parameterized by multi-layer perceptron (MLP) layers, to learn to estimate oriented normals from point clouds in an end-to-end manner. The signed hyper surfaces are implicitly learned in a high-dimensional feature space where the local and global information is aggregated. Specifically, we introduce a patch encoding module and a shape encoding module to encode a 3D point cloud into a local latent code and a global latent code, respectively. Then, an attention-weighted normal prediction module is proposed as a decoder, which takes the local and global latent codes as input to predict oriented normals. Experimental results show that our SHS-Net outperforms the state-of-the-art methods in both unoriented and oriented normal estimation on the widely used benchmarks.

Xinhai Liu, Zhizhong Han, Sanghuk Lee et al.

Learning and selecting important points on a point cloud is crucial for point cloud understanding in various applications. Most of early methods selected the important points on 3D shapes by analyzing the intrinsic geometric properties of every single shape, which fails to capture the importance of points that distinguishes a shape from objects of other classes, i.e., the distinction of points. To address this problem, we propose D-Net (Distinctive Network) to learn for distinctive point clouds based on a self-attentive point searching and a learnable feature fusion. Specifically, in the self-attentive point searching, we first learn the distinction score for each point to reveal the distinction distribution of the point cloud. After ranking the learned distinction scores, we group a point cloud into a high distinctive point set and a low distinctive one to enrich the fine-grained point cloud structure. To generate a compact feature representation for each distinctive point set, a stacked self-gated convolution is proposed to extract the distinctive features. Finally, we further introduce a learnable feature fusion mechanism to aggregate multiple distinctive features into a global point cloud representation in a channel-wise aggregation manner. The results also show that the learned distinction distribution of a point cloud is highly consistent with objects of the same class and different from objects of other classes. Extensive experiments on public datasets, including ModelNet and ShapeNet part dataset, demonstrate the ability to learn for distinctive point clouds, which helps to achieve the state-of-the-art performance in some shape understanding applications.

Zhizhong Han, Xiyang Wang, Yu-Shen Liu et al.

Unsupervised learning of global features for 3D shape analysis is an important research challenge because it avoids manual effort for supervised information collection. In this paper, we propose a view-based deep learning model called Hierarchical View Predictor (HVP) to learn 3D shape features from unordered views in an unsupervised manner. To mine highly discriminative information from unordered views, HVP performs a novel hierarchical view prediction over a view pair, and aggregates the knowledge learned from the predictions in all view pairs into a global feature. In a view pair, we pose hierarchical view prediction as the task of hierarchically predicting a set of image patches in a current view from its complementary set of patches, and in addition, completing the current view and its opposite from any one of the two sets of patches. Hierarchical prediction, in patches to patches, patches to view and view to view, facilitates HVP to effectively learn the structure of 3D shapes from the correlation between patches in the same view and the correlation between a pair of complementary views. In addition, the employed implicit aggregation over all view pairs enables HVP to learn global features from unordered views. Our results show that HVP can outperform state-of-the-art methods under large-scale 3D shape benchmarks in shape classification and retrieval.

Xin Wen, Zhizhong Han, Yan-Pei Cao et al.

In this paper, we present a novel unpaired point cloud completion network, named Cycle4Completion, to infer the complete geometries from a partial 3D object. Previous unpaired completion methods merely focus on the learning of geometric correspondence from incomplete shapes to complete shapes, and ignore the learning in the reverse direction, which makes them suffer from low completion accuracy due to the limited 3D shape understanding ability. To address this problem, we propose two simultaneous cycle transformations between the latent spaces of complete shapes and incomplete ones. The insight of cycle transformation is to promote networks to understand 3D shapes by learning to generate complete or incomplete shapes from their complementary ones. Specifically, the first cycle transforms shapes from incomplete domain to complete domain, and then projects them back to the incomplete domain. This process learns the geometric characteristic of complete shapes, and maintains the shape consistency between the complete prediction and the incomplete input. Similarly, the inverse cycle transformation starts from complete domain to incomplete domain, and goes back to complete domain to learn the characteristic of incomplete shapes. We provide a comprehensive evaluation in experiments, which shows that our model with the learned bidirectional geometry correspondence outperforms state-of-the-art unpaired completion methods.

Xin Wen, Peng Xiang, Zhizhong Han et al.

The task of point cloud completion aims to predict the missing part for an incomplete 3D shape. A widely used strategy is to generate a complete point cloud from the incomplete one. However, the unordered nature of point clouds will degrade the generation of high-quality 3D shapes, as the detailed topology and structure of discrete points are hard to be captured by the generative process only using a latent code. In this paper, we address the above problem by reconsidering the completion task from a new perspective, where we formulate the prediction as a point cloud deformation process. Specifically, we design a novel neural network, named PMP-Net, to mimic the behavior of an earth mover. It moves each point of the incomplete input to complete the point cloud, where the total distance of point moving paths (PMP) should be shortest. Therefore, PMP-Net predicts a unique point moving path for each point according to the constraint of total point moving distances. As a result, the network learns a strict and unique correspondence on point-level, which can capture the detailed topology and structure relationships between the incomplete shape and the complete target, and thus improves the quality of the predicted complete shape. We conduct comprehensive experiments on Completion3D and PCN datasets, which demonstrate our advantages over the state-of-the-art point cloud completion methods.

Baorui Ma, Zhizhong Han, Yu-Shen Liu et al.

Reconstructing continuous surfaces from 3D point clouds is a fundamental operation in 3D geometry processing. Several recent state-of-the-art methods address this problem using neural networks to learn signed distance functions (SDFs). In this paper, we introduce \textit{Neural-Pull}, a new approach that is simple and leads to high quality SDFs. Specifically, we train a neural network to pull query 3D locations to their closest points on the surface using the predicted signed distance values and the gradient at the query locations, both of which are computed by the network itself. The pulling operation moves each query location with a stride given by the distance predicted by the network. Based on the sign of the distance, this may move the query location along or against the direction of the gradient of the SDF. This is a differentiable operation that allows us to update the signed distance value and the gradient simultaneously during training. Our outperforming results under widely used benchmarks demonstrate that we can learn SDFs more accurately and flexibly for surface reconstruction and single image reconstruction than the state-of-the-art methods.

Zhizhong Han, Chao Chen, Yu-Shen Liu et al.

Differentiable renderers have been used successfully for unsupervised 3D structure learning from 2D images because they can bridge the gap between 3D and 2D. To optimize 3D shape parameters, current renderers rely on pixel-wise losses between rendered images of 3D reconstructions and ground truth images from corresponding viewpoints. Hence they require interpolation of the recovered 3D structure at each pixel, visibility handling, and optionally evaluating a shading model. In contrast, here we propose a Differentiable Renderer Without Rendering (DRWR) that omits these steps. DRWR only relies on a simple but effective loss that evaluates how well the projections of reconstructed 3D point clouds cover the ground truth object silhouette. Specifically, DRWR employs a smooth silhouette loss to pull the projection of each individual 3D point inside the object silhouette, and a structure-aware repulsion loss to push each pair of projections that fall inside the silhouette far away from each other. Although we omit surface interpolation, visibility handling, and shading, our results demonstrate that DRWR achieves state-of-the-art accuracies under widely used benchmarks, outperforming previous methods both qualitatively and quantitatively. In addition, our training times are significantly lower due to the simplicity of DRWR.

Xin Wen, Tianyang Li, Zhizhong Han et al.

Point cloud completion aims to infer the complete geometries for missing regions of 3D objects from incomplete ones. Previous methods usually predict the complete point cloud based on the global shape representation extracted from the incomplete input. However, the global representation often suffers from the information loss of structure details on local regions of incomplete point cloud. To address this problem, we propose Skip-Attention Network (SA-Net) for 3D point cloud completion. Our main contributions lie in the following two-folds. First, we propose a skip-attention mechanism to effectively exploit the local structure details of incomplete point clouds during the inference of missing parts. The skip-attention mechanism selectively conveys geometric information from the local regions of incomplete point clouds for the generation of complete ones at different resolutions, where the skip-attention reveals the completion process in an interpretable way. Second, in order to fully utilize the selected geometric information encoded by skip-attention mechanism at different resolutions, we propose a novel structure-preserving decoder with hierarchical folding for complete shape generation. The hierarchical folding preserves the structure of complete point cloud generated in upper layer by progressively detailing the local regions, using the skip-attentioned geometry at the same resolution. We conduct comprehensive experiments on ShapeNet and KITTI datasets, which demonstrate that the proposed SA-Net outperforms the state-of-the-art point cloud completion methods.

Xinhai Liu, Zhizhong Han, Fangzhou Hong et al.

Learning discriminative feature directly on point clouds is still challenging in the understanding of 3D shapes. Recent methods usually partition point clouds into local region sets, and then extract the local region features with fixed-size CNN or MLP, and finally aggregate all individual local features into a global feature using simple max pooling. However, due to the irregularity and sparsity in sampled point clouds, it is hard to encode the fine-grained geometry of local regions and their spatial relationships when only using the fixed-size filters and individual local feature integration, which limit the ability to learn discriminative features. To address this issue, we present a novel Local-Region-Context Network (LRC-Net), to learn discriminative features on point clouds by encoding the fine-grained contexts inside and among local regions simultaneously. LRC-Net consists of two main modules. The first module, named intra-region context encoding, is designed for capturing the geometric correlation inside each local region by novel variable-size convolution filter. The second module, named inter-region context encoding, is proposed for integrating the spatial relationships among local regions based on spatial similarity measures. Experimental results show that LRC-Net is competitive with state-of-the-art methods in shape classification and shape segmentation applications.

Zhizhong Han, Guanhui Qiao, Yu-Shen Liu et al.

Structure learning for 3D shapes is vital for 3D computer vision. State-of-the-art methods show promising results by representing shapes using implicit functions in 3D that are learned using discriminative neural networks. However, learning implicit functions requires dense and irregular sampling in 3D space, which also makes the sampling methods affect the accuracy of shape reconstruction during test. To avoid dense and irregular sampling in 3D, we propose to represent shapes using 2D functions, where the output of the function at each 2D location is a sequence of line segments inside the shape. Our approach leverages the power of functional representations, but without the disadvantage of 3D sampling. Specifically, we use a voxel tubelization to represent a voxel grid as a set of tubes along any one of the X, Y, or Z axes. Each tube can be indexed by its 2D coordinates on the plane spanned by the other two axes. We further simplify each tube into a sequence of occupancy segments. Each occupancy segment consists of successive voxels occupied by the shape, which leads to a simple representation of its 1D start and end location. Given the 2D coordinates of the tube and a shape feature as condition, this representation enables us to learn 3D shape structures by sequentially predicting the start and end locations of each occupancy segment in the tube. We implement this approach using a Seq2Seq model with attention, called SeqXY2SeqZ, which learns the mapping from a sequence of 2D coordinates along two arbitrary axes to a sequence of 1D locations along the third axis. SeqXY2SeqZ not only benefits from the regularity of voxel grids in training and testing, but also achieves high memory efficiency. Our experiments show that SeqXY2SeqZ outperforms the state-ofthe-art methods under widely used benchmarks.

Yue Jiang, Dantong Ji, Zhizhong Han et al.

We propose SDFDiff, a novel approach for image-based shape optimization using differentiable rendering of 3D shapes represented by signed distance functions (SDFs). Compared to other representations, SDFs have the advantage that they can represent shapes with arbitrary topology, and that they guarantee watertight surfaces. We apply our approach to the problem of multi-view 3D reconstruction, where we achieve high reconstruction quality and can capture complex topology of 3D objects. In addition, we employ a multi-resolution strategy to obtain a robust optimization algorithm. We further demonstrate that our SDF-based differentiable renderer can be integrated with deep learning models, which opens up options for learning approaches on 3D objects without 3D supervision. In particular, we apply our method to single-view 3D reconstruction and achieve state-of-the-art results.

Tao Hu, Zhizhong Han, Matthias Zwicker

3D shape completion is important to enable machines to perceive the complete geometry of objects from partial observations. To address this problem, view-based methods have been presented. These methods represent shapes as multiple depth images, which can be back-projected to yield corresponding 3D point clouds, and they perform shape completion by learning to complete each depth image using neural networks. While view-based methods lead to state-of-the-art results, they currently do not enforce geometric consistency among the completed views during the inference stage. To resolve this issue, we propose a multi-view consistent inference technique for 3D shape completion, which we express as an energy minimization problem including a data term and a regularization term. We formulate the regularization term as a consistency loss that encourages geometric consistency among multiple views, while the data term guarantees that the optimized views do not drift away too much from a learned shape descriptor. Experimental results demonstrate that our method completes shapes more accurately than previous techniques.

Xin Wen, Zhizhong Han, Xinhai Liu et al.

Learning discriminative shape representation directly on point clouds is still challenging in 3D shape analysis and understanding. Recent studies usually involve three steps: first splitting a point cloud into some local regions, then extracting corresponding feature of each local region, and finally aggregating all individual local region features into a global feature as shape representation using simple max pooling. However, such pooling-based feature aggregation methods do not adequately take the spatial relationships between local regions into account, which greatly limits the ability to learn discriminative shape representation. To address this issue, we propose a novel deep learning network, named Point2SpatialCapsule, for aggregating features and spatial relationships of local regions on point clouds, which aims to learn more discriminative shape representation. Compared with traditional max-pooling based feature aggregation networks, Point2SpatialCapsule can explicitly learn not only geometric features of local regions but also spatial relationships among them. It consists of two modules. To resolve the disorder problem of local regions, the first module, named geometric feature aggregation, is designed to aggregate the local region features into the learnable cluster centers, which explicitly encodes the spatial locations from the original 3D space. The second module, named spatial relationship aggregation, is proposed for further aggregating clustered features and the spatial relationships among them in the feature space using the spatial-aware capsules developed in this paper. Compared to the previous capsule network based methods, the feature routing on the spatial-aware capsules can learn more discriminative spatial relationships among local regions for point clouds, which establishes a direct mapping between log priors and the spatial locations through feature clusters.

Xinhai Liu, Zhizhong Han, Xin Wen et al.

Auto-encoder is an important architecture to understand point clouds in an encoding and decoding procedure of self reconstruction. Current auto-encoder mainly focuses on the learning of global structure by global shape reconstruction, while ignoring the learning of local structures. To resolve this issue, we propose Local-to-Global auto-encoder (L2G-AE) to simultaneously learn the local and global structure of point clouds by local to global reconstruction. Specifically, L2G-AE employs an encoder to encode the geometry information of multiple scales in a local region at the same time. In addition, we introduce a novel hierarchical self-attention mechanism to highlight the important points, scales and regions at different levels in the information aggregation of the encoder. Simultaneously, L2G-AE employs a recurrent neural network (RNN) as decoder to reconstruct a sequence of scales in a local region, based on which the global point cloud is incrementally reconstructed. Our outperforming results in shape classification, retrieval and upsampling show that L2G-AE can understand point clouds better than state-of-the-art methods.

Zhizhong Han, Chao Chen, Yu-Shen Liu et al.

3D shape captioning is a challenging application in 3D shape understanding. Captions from recent multi-view based methods reveal that they cannot capture part-level characteristics of 3D shapes. This leads to a lack of detailed part-level description in captions, which human tend to focus on. To resolve this issue, we propose ShapeCaptioner, a generative caption network, to perform 3D shape captioning from semantic parts detected in multiple views. Our novelty lies in learning the knowledge of part detection in multiple views from 3D shape segmentations and transferring this knowledge to facilitate learning the mapping from 3D shapes to sentences. Specifically, ShapeCaptioner aggregates the parts detected in multiple colored views using our novel part class specific aggregation to represent a 3D shape, and then, employs a sequence to sequence model to generate the caption. Our outperforming results show that ShapeCaptioner can learn 3D shape features with more detailed part characteristics to facilitate better 3D shape captioning than previous work.

Zhizhong Han, Xiyang Wang, Yu-Shen Liu et al.

Unsupervised feature learning for point clouds has been vital for large-scale point cloud understanding. Recent deep learning based methods depend on learning global geometry from self-reconstruction. However, these methods are still suffering from ineffective learning of local geometry, which significantly limits the discriminability of learned features. To resolve this issue, we propose MAP-VAE to enable the learning of global and local geometry by jointly leveraging global and local self-supervision. To enable effective local self-supervision, we introduce multi-angle analysis for point clouds. In a multi-angle scenario, we first split a point cloud into a front half and a back half from each angle, and then, train MAP-VAE to learn to predict a back half sequence from the corresponding front half sequence. MAP-VAE performs this half-to-half prediction using RNN to simultaneously learn each local geometry and the spatial relationship among them. In addition, MAP-VAE also learns global geometry via self-reconstruction, where we employ a variational constraint to facilitate novel shape generation. The outperforming results in four shape analysis tasks show that MAP-VAE can learn more discriminative global or local features than the state-of-the-art methods.

Zhizhong Han, Xinhai Liu, Yu-Shen Liu et al.

Deep learning has achieved remarkable results in 3D shape analysis by learning global shape features from the pixel-level over multiple views. Previous methods, however, compute low-level features for entire views without considering part-level information. In contrast, we propose a deep neural network, called Parts4Feature, to learn 3D global features from part-level information in multiple views. We introduce a novel definition of generally semantic parts, which Parts4Feature learns to detect in multiple views from different 3D shape segmentation benchmarks. A key idea of our architecture is that it transfers the ability to detect semantically meaningful parts in multiple views to learn 3D global features. Parts4Feature achieves this by combining a local part detection branch and a global feature learning branch with a shared region proposal module. The global feature learning branch aggregates the detected parts in terms of learned part patterns with a novel multi-attention mechanism, while the region proposal module enables locally and globally discriminative information to be promoted by each other. We demonstrate that Parts4Feature outperforms the state-of-the-art under three large-scale 3D shape benchmarks.

Zhizhong Han, Xiyang Wang, Chi-Man Vong et al.

Learning global features by aggregating information over multiple views has been shown to be effective for 3D shape analysis. For view aggregation in deep learning models, pooling has been applied extensively. However, pooling leads to a loss of the content within views, and the spatial relationship among views, which limits the discriminability of learned features. We propose 3DViewGraph to resolve this issue, which learns 3D global features by more effectively aggregating unordered views with attention. Specifically, unordered views taken around a shape are regarded as view nodes on a view graph. 3DViewGraph first learns a novel latent semantic mapping to project low-level view features into meaningful latent semantic embeddings in a lower dimensional space, which is spanned by latent semantic patterns. Then, the content and spatial information of each pair of view nodes are encoded by a novel spatial pattern correlation, where the correlation is computed among latent semantic patterns. Finally, all spatial pattern correlations are integrated with attention weights learned by a novel attention mechanism. This further increases the discriminability of learned features by highlighting the unordered view nodes with distinctive characteristics and depressing the ones with appearance ambiguity. We show that 3DViewGraph outperforms state-of-the-art methods under three large-scale benchmarks.

Xin Wen, Zhizhong Han, Xinyu Yin et al.

Cross-modal retrieval aims to retrieve relevant data across different modalities (e.g., texts vs. images). The common strategy is to apply element-wise constraints between manually labeled pair-wise items to guide the generators to learn the semantic relationships between the modalities, so that the similar items can be projected close to each other in the common representation subspace. However, such constraints often fail to preserve the semantic structure between unpaired but semantically similar items (e.g. the unpaired items with the same class label are more similar than items with different labels). To address the above problem, we propose a novel cross-modal similarity transferring (CMST) method to learn and preserve the semantic relationships between unpaired items in an unsupervised way. The key idea is to learn the quantitative similarities in single-modal representation subspace, and then transfer them to the common representation subspace to establish the semantic relationships between unpaired items across modalities. Experiments show that our method outperforms the state-of-the-art approaches both in the class-based and pair-based retrieval tasks.

Zhizhong Han, Mingyang Shang, Xiyang Wang et al.

A recent method employs 3D voxels to represent 3D shapes, but this limits the approach to low resolutions due to the computational cost caused by the cubic complexity of 3D voxels. Hence the method suffers from a lack of detailed geometry. To resolve this issue, we propose Y^2Seq2Seq, a view-based model, to learn cross-modal representations by joint reconstruction and prediction of view and word sequences. Specifically, the network architecture of Y^2Seq2Seq bridges the semantic meaning embedded in the two modalities by two coupled `Y' like sequence-to-sequence (Seq2Seq) structures. In addition, our novel hierarchical constraints further increase the discriminability of the cross-modal representations by employing more detailed discriminative information. Experimental results on cross-modal retrieval and 3D shape captioning show that Y^2Seq2Seq outperforms the state-of-the-art methods.

Zhizhong Han, Mingyang Shang, Yu-Shen Liu et al.

In this paper we present a novel unsupervised representation learning approach for 3D shapes, which is an important research challenge as it avoids the manual effort required for collecting supervised data. Our method trains an RNN-based neural network architecture to solve multiple view inter-prediction tasks for each shape. Given several nearby views of a shape, we define view inter-prediction as the task of predicting the center view between the input views, and reconstructing the input views in a low-level feature space. The key idea of our approach is to implement the shape representation as a shape-specific global memory that is shared between all local view inter-predictions for each shape. Intuitively, this memory enables the system to aggregate information that is useful to better solve the view inter-prediction tasks for each shape, and to leverage the memory as a view-independent shape representation. Our approach obtains the best results using a combination of L_2 and adversarial losses for the view inter-prediction task. We show that VIP-GAN outperforms state-of-the-art methods in unsupervised 3D feature learning on three large scale 3D shape benchmarks.

Xinhai Liu, Zhizhong Han, Yu-Shen Liu et al.

Exploring contextual information in the local region is important for shape understanding and analysis. Existing studies often employ hand-crafted or explicit ways to encode contextual information of local regions. However, it is hard to capture fine-grained contextual information in hand-crafted or explicit manners, such as the correlation between different areas in a local region, which limits the discriminative ability of learned features. To resolve this issue, we propose a novel deep learning model for 3D point clouds, named Point2Sequence, to learn 3D shape features by capturing fine-grained contextual information in a novel implicit way. Point2Sequence employs a novel sequence learning model for point clouds to capture the correlations by aggregating multi-scale areas of each local region with attention. Specifically, Point2Sequence first learns the feature of each area scale in a local region. Then, it captures the correlation between area scales in the process of aggregating all area scales using a recurrent neural network (RNN) based encoder-decoder structure, where an attention mechanism is proposed to highlight the importance of different area scales. Experimental results show that Point2Sequence achieves state-of-the-art performance in shape classification and segmentation tasks.