Yi Fang

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.

Yan Luo, Congcong Wen, Min Shi et al.

We present a comprehensive theoretical framework analyzing the relationship between data distributions and fairness guarantees in equitable deep learning. Our work establishes novel theoretical bounds that explicitly account for data distribution heterogeneity across demographic groups, while introducing a formal analysis framework that minimizes expected loss differences across these groups. We derive comprehensive theoretical bounds for fairness errors and convergence rates, and characterize how distributional differences between groups affect the fundamental trade-off between fairness and accuracy. Through extensive experiments on diverse datasets, including FairVision (ophthalmology), CheXpert (chest X-rays), HAM10000 (dermatology), and FairFace (facial recognition), we validate our theoretical findings and demonstrate that differences in feature distributions across demographic groups significantly impact model fairness, with performance disparities particularly pronounced in racial categories. The theoretical bounds we derive crroborate these empirical observations, providing insights into the fundamental limits of achieving fairness in deep learning models when faced with heterogeneous data distributions. This work advances our understanding of fairness in AI-based diagnosis systems and provides a theoretical foundation for developing more equitable algorithms. The code for analysis is publicly available via \url{https://github.com/Harvard-Ophthalmology-AI-Lab/fairness_guarantees}.

Lingjing Wang, Jianchun Chen, Xiang Li et al.

Point set registration is defined as a process to determine the spatial transformation from the source point set to the target one. Existing methods often iteratively search for the optimal geometric transformation to register a given pair of point sets, driven by minimizing a predefined alignment loss function. In contrast, the proposed point registration neural network (PR-Net) actively learns the registration pattern as a parametric function from a training dataset, consequently predict the desired geometric transformation to align a pair of point sets. PR-Net can transfer the learned knowledge (i.e. registration pattern) from registering training pairs to testing ones without additional iterative optimization. Specifically, in this paper, we develop novel techniques to learn shape descriptors from point sets that help formulate a clear correlation between source and target point sets. With the defined correlation, PR-Net tends to predict the transformation so that the source and target point sets can be statistically aligned, which in turn leads to an optimal spatial geometric registration. PR-Net achieves robust and superior performance for non-rigid registration of point sets, even in presence of Gaussian noise, outliers, and missing points, but requires much less time for registering large number of pairs. More importantly, for a new pair of point sets, PR-Net is able to directly predict the desired transformation using the learned model without repetitive iterative optimization routine. Our code is available at https://github.com/Lingjing324/PR-Net.

Congcong Wen, Yisiyuan Huang, Hao Huang et al.

Object navigation is crucial for robots, but traditional methods require substantial training data and cannot be generalized to unknown environments. Zero-shot object navigation (ZSON) aims to address this challenge, allowing robots to interact with unknown objects without specific training data. Language-driven zero-shot object navigation (L-ZSON) is an extension of ZSON that incorporates natural language instructions to guide robot navigation and interaction with objects. In this paper, we propose a novel Vision Language model with a Tree-of-thought Network (VLTNet) for L-ZSON. VLTNet comprises four main modules: vision language model understanding, semantic mapping, tree-of-thought reasoning and exploration, and goal identification. Among these modules, Tree-of-Thought (ToT) reasoning and exploration module serves as a core component, innovatively using the ToT reasoning framework for navigation frontier selection during robot exploration. Compared to conventional frontier selection without reasoning, navigation using ToT reasoning involves multi-path reasoning processes and backtracking when necessary, enabling globally informed decision-making with higher accuracy. Experimental results on PASTURE and RoboTHOR benchmarks demonstrate the outstanding performance of our model in LZSON, particularly in scenarios involving complex natural language as target instructions.

Ting-Ruen Wei, Haowei Liu, Xuyang Wu et al.

Recent progress in large language models (LLM) found chain-of-thought prompting strategies to improve the reasoning ability of LLMs by encouraging problem solving through multiple steps. Therefore, subsequent research aimed to integrate the multi-step reasoning process into the LLM itself through process rewards as feedback and achieved improvements over prompting strategies. Due to the cost of step-level annotation, some turn to outcome rewards as feedback. Aside from these training-based approaches, training-free techniques leverage frozen LLMs or external tools for feedback at each step to enhance the reasoning process. With the abundance of work in mathematics due to its logical nature, we present a survey of strategies utilizing feedback at the step and outcome levels to enhance multi-step math reasoning for LLMs. As multi-step reasoning emerges a crucial component in scaling LLMs, we hope to establish its foundation for easier understanding and empower further research.

Hao Huang, Shuaihang Yuan, Yu Hao et al.

A large-scale vision and language model that has been pretrained on massive data encodes visual and linguistic prior, which makes it easier to generate images and language that are more natural and realistic. Despite this, there is still a significant domain gap between the modalities of vision and language, especially when training data is scarce in few-shot settings, where only very limited data are available for training. In order to mitigate this issue, a multi-modal meta-learning framework has been proposed to bridge the gap between two frozen pretrained large vision and language models by introducing a tunable prompt connecting these two large models. For few-shot image captioning, the existing multi-model meta-learning framework utilizes a one-step prompting scheme to accumulate the visual features of input images to guide the language model, which struggles to generate accurate image descriptions with only a few training samples. Instead, we propose a chain-of-thought (CoT) meta-learning scheme as a multi-step image captioning procedure to better imitate how humans describe images. In addition, we further propose to learn different meta-parameters of the model corresponding to each CoT step in distinct subspaces to avoid interference. We evaluated our method on three commonly used image captioning datasets, i.e., MSCOCO, Flickr8k, and Flickr30k, under few-shot settings. The results of our experiments indicate that our chain-of-thought subspace meta-learning strategy is superior to the baselines in terms of performance across different datasets measured by different metrics.

Halil Utku Unlu, Shuaihang Yuan, Congcong Wen et al.

We introduce an innovative approach to advancing semantic understanding in zero-shot object goal navigation (ZS-OGN), enhancing the autonomy of robots in unfamiliar environments. Traditional reliance on labeled data has been a limitation for robotic adaptability, which we address by employing a dual-component framework that integrates a GLIP Vision Language Model for initial detection and an InstructionBLIP model for validation. This combination not only refines object and environmental recognition but also fortifies the semantic interpretation, pivotal for navigational decision-making. Our method, rigorously tested in both simulated and real-world settings, exhibits marked improvements in navigation precision and reliability.

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.

Yu Hao, Yi Fang

Though deep learning methods have shown great success in 3D point cloud part segmentation, they generally rely on a large volume of labeled training data, which makes the model suffer from unsatisfied generalization abilities to unseen classes with limited data. To address this problem, we present a novel meta-learning strategy that regards the 3D shape segmentation function as a task. By training over a number of 3D part segmentation tasks, our method is capable to learn the prior over the respective 3D segmentation function space which leads to an optimal model that is rapidly adapting to new part segmentation tasks. To implement our meta-learning strategy, we propose two novel modules: meta part segmentation learner and part segmentation learner. During the training process, the part segmentation learner is trained to complete a specific part segmentation task in the few-shot scenario. In the meantime, the meta part segmentation learner is trained to capture the prior from multiple similar part segmentation tasks. Based on the learned information of task distribution, our meta part segmentation learner is able to dynamically update the part segmentation learner with optimal parameters which enable our part segmentation learner to rapidly adapt and have great generalization ability on new part segmentation tasks. We demonstrate that our model achieves superior part segmentation performance with the few-shot setting on the widely used dataset: ShapeNet.

Yu Hao, Yi Fang

This paper concerns the research problem of point cloud registration to find the rigid transformation to optimally align the source point set with the target one. Learning robust point cloud registration models with deep neural networks has emerged as a powerful paradigm, offering promising performance in predicting the global geometric transformation for a pair of point sets. Existing methods first leverage an encoder to regress the global shape descriptor, which is then decoded into a shape-conditioned transformation via concatenation-based conditioning. However, different regions of a 3D shape vary in their geometric structures which makes it more sense that we have a region-conditioned transformation instead of the shape-conditioned one. In this paper, we define our 3D registration function through the introduction of a new design of 3D region partition module that is able to divide the input shape to different regions with a self-supervised 3D shape reconstruction loss without the need for ground truth labels. We further propose the 3D shape transformer module to efficiently and effectively capture short- and long-range geometric dependencies for regions on the 3D shape Consequently, the region-aware decoder module is proposed to predict the transformations for different regions respectively. The global geometric transformation from the source point set to the target one is then formed by the weighted fusion of region-aware transformation. Compared to the state-of-the-art approaches, our experiments show that our 3D-URRT achieves superior registration performance over various benchmark datasets (e.g. ModelNet40).

Mengxi Wu, Hao Huang, Yi Fang

With the popularity of 3D sensors in self-driving and other robotics applications, extensive research has focused on designing novel neural network architectures for accurate 3D point cloud completion. However, unlike in point cloud classification and reconstruction, the role of adversarial samples in3D point cloud completion has seldom been explored. In this work, we show that training with adversarial samples can improve the performance of neural networks on 3D point cloud completion tasks. We propose a novel approach to generate adversarial samples that benefit both the performance of clean and adversarial samples. In contrast to the PGD-k attack, our method generates adversarial samples that keep the geometric features in clean samples and contain few outliers. In particular, we use principal directions to constrain the adversarial perturbations for each input point. The gradient components in the mean direction of principal directions are taken as adversarial perturbations. In addition, we also investigate the effect of using the minimum curvature direction. Besides, we adopt attack strength accumulation and auxiliary Batch Normalization layers method to speed up the training process and alleviate the distribution mismatch between clean and adversarial samples. Experimental results show that training with the adversarial samples crafted by our method effectively enhances the performance of PCN on the ShapeNet dataset.

Haoyu Fang, Jing Zhu, Yi Fang

Lane segmentation is a challenging issue in autonomous driving system designing because lane marks show weak textural consistency due to occlusion or extreme illumination but strong geometric continuity in traffic images, from which general convolution neural networks (CNNs) are not capable of learning semantic objects. To empower conventional CNNs in learning geometric clues of lanes, we propose a deep network named ContinuityLearner to better learn geometric prior within lane. Specifically, our proposed CNN-based paradigm involves a novel Context-encoding image feature learning network to generate class-dependent image feature maps and a new encoding layer to exploit the geometric continuity feature representation by fusing both spatial and visual information of lane together. The ContinuityLearner, performing on the geometric continuity feature of lanes, is trained to directly predict the lane in traffic scenarios with integrated and continuous instance semantic. The experimental results on the CULane dataset and the Tusimple benchmark demonstrate that our ContinuityLearner has superior performance over other state-of-the-art techniques in lane segmentation.

Xiang Li, Lingjing Wang, Yi Fang

Recent research has seen numerous supervised learning-based methods for 3D shape segmentation and remarkable performance has been achieved on various benchmark datasets. These supervised methods require a large amount of annotated data to train deep neural networks to ensure the generalization ability on the unseen test set. In this paper, we introduce a meta-learning-based method for few-shot 3D shape segmentation where only a few labeled samples are provided for the unseen classes. To achieve this, we treat the shape segmentation as a point labeling problem in the metric space. Specifically, we first design a meta-metric learner to transform input shapes into embedding space and our model learns to learn a proper metric space for each object class based on point embeddings. Then, for each class, we design a metric learner to extract part-specific prototype representations from a few support shapes and our model performs per-point segmentation over the query shapes by matching each point to its nearest prototype in the learned metric space. A metric-based loss function is used to dynamically modify distances between point embeddings thus maximizes in-part similarity while minimizing inter-part similarity. A dual segmentation branch is adopted to make full use of the support information and implicitly encourages consistency between the support and query prototypes. We demonstrate the superior performance of our proposed on the ShapeNet part dataset under the few-shot scenario, compared with well-established baseline and state-of-the-art semi-supervised methods.

Hao Huang, Boulbaba Ben Amor, Xichan Lin et al.

Analyzing the structure of proteins is a key part of understanding their functions and thus their role in biology at the molecular level. In addition, design new proteins in a methodical way is a major engineering challenge. In this work, we introduce a joint geometric-neural networks approach for comparing, deforming and generating 3D protein structures. Viewing protein structures as 3D open curves, we adopt the Square Root Velocity Function (SRVF) representation and leverage its suitable geometric properties along with Deep Residual Networks (ResNets) for a joint registration and comparison. Our ResNets handle better large protein deformations while being more computationally efficient. On top of the mathematical framework, we further design a Geometric Variational Auto-Encoder (G-VAE), that once trained, maps original, previously unseen structures, into a low-dimensional (latent) hyper-sphere. Motivated by the spherical structure of the pre-shape space, we naturally adopt the von Mises-Fisher (vMF) distribution to model our hidden variables. We test the effectiveness of our models by generating novel protein structures and predicting completions of corrupted protein structures. Experimental results show that our method is able to generate plausible structures, different from the structures in the training data.

Hao Huang, Boulbaba Ben Amor, Xichan Lin et al.

Non-linear (large) time warping is a challenging source of nuisance in time-series analysis. In this paper, we propose a novel diffeomorphic temporal transformer network for both pairwise and joint time-series alignment. Our ResNet-TW (Deep Residual Network for Time Warping) tackles the alignment problem by compositing a flow of incremental diffeomorphic mappings. Governed by the flow equation, our Residual Network (ResNet) builds smooth, fluid and regular flows of velocity fields and consequently generates smooth and invertible transformations (i.e. diffeomorphic warping functions). Inspired by the elegant Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework, the final transformation is built by the flow of time-dependent vector fields which are none other than the building blocks of our Residual Network. The latter is naturally viewed as an Eulerian discretization schema of the flow equation (an ODE). Once trained, our ResNet-TW aligns unseen data by a single inexpensive forward pass. As we show in experiments on both univariate (84 datasets from UCR archive) and multivariate time-series (MSR Action-3D, Florence-3D and MSR Daily Activity), ResNet-TW achieves competitive performance in joint alignment and classification.

Lingjing Wang, Yu Hao, Xiang Li et al.

Deep learning-based point cloud registration models are often generalized from extensive training over a large volume of data to learn the ability to predict the desired geometric transformation to register 3D point clouds. In this paper, we propose a meta-learning based 3D registration model, named 3D Meta-Registration, that is capable of rapidly adapting and well generalizing to new 3D registration tasks for unseen 3D point clouds. Our 3D Meta-Registration gains a competitive advantage by training over a variety of 3D registration tasks, which leads to an optimized model for the best performance on the distribution of registration tasks including potentially unseen tasks. Specifically, the proposed 3D Meta-Registration model consists of two modules: 3D registration learner and 3D registration meta-learner. During the training, the 3D registration learner is trained to complete a specific registration task aiming to determine the desired geometric transformation that aligns the source point cloud with the target one. In the meantime, the 3D registration meta-learner is trained to provide the optimal parameters to update the 3D registration learner based on the learned task distribution. After training, the 3D registration meta-learner, which is learned with the optimized coverage of distribution of 3D registration tasks, is able to dynamically update 3D registration learners with desired parameters to rapidly adapt to new registration tasks. We tested our model on synthesized dataset ModelNet and FlyingThings3D, as well as real-world dataset KITTI. Experimental results demonstrate that 3D Meta-Registration achieves superior performance over other previous techniques (e.g. FlowNet3D).

Hao Huang, Lingjing Wang, Xiang Li et al.

Point signature, a representation describing the structural neighborhood of a point in 3D shapes, can be applied to establish correspondences between points in 3D shapes. Conventional methods apply a weight-sharing network, e.g., any kind of graph neural networks, across all neighborhoods to directly generate point signatures and gain the generalization ability by extensive training over a large amount of training samples from scratch. However, these methods lack the flexibility in rapidly adapting to unseen neighborhood structures and thus generalizes poorly on new point sets. In this paper, we propose a novel meta-learning based 3D point signature model, named 3Dmetapointsignature (MEPS) network, that is capable of learning robust point signatures in 3D shapes. By regarding each point signature learning process as a task, our method obtains an optimized model over the best performance on the distribution of all tasks, generating reliable signatures for new tasks, i.e., signatures of unseen point neighborhoods. Specifically, the MEPS consists of two modules: a base signature learner and a meta signature learner. During training, the base-learner is trained to perform specific signature learning tasks. In the meantime, the meta-learner is trained to update the base-learner with optimal parameters. During testing, the meta-learner that is learned with the distribution of all tasks can adaptively change parameters of the base-learner, accommodating to unseen local neighborhoods. We evaluate the MEPS model on two datasets, e.g., FAUST and TOSCA, for dense 3Dshape correspondence. Experimental results demonstrate that our method not only gains significant improvements over the baseline model and achieves state-of-the-art results, but also is capable of handling unseen 3D shapes.

Lingjing Wang, Xiang Li, Yi Fang

Point cloud registration is the process of aligning a pair of point sets via searching for a geometric transformation. Unlike classical optimization-based methods, recent learning-based methods leverage the power of deep learning for registering a pair of point sets. In this paper, we propose to develop a novel model that organically integrates the optimization to learning, aiming to address the technical challenges in 3D registration. More specifically, in addition to the deep transformation decoding network, our framework introduce an optimizable deep \underline{S}patial \underline{C}orrelation \underline{R}epresentation (SCR) feature. The SCR feature and weights of the transformation decoder network are jointly updated towards the minimization of an unsupervised alignment loss. We further propose an adaptive Chamfer loss for aligning partial shapes. To verify the performance of our proposed method, we conducted extensive experiments on the ModelNet40 dataset. The results demonstrate that our method achieves significantly better performance than the previous state-of-the-art approaches in the full/partial point set registration task.

Xiang Li, Lingjing Wang, Yi Fang

We propose a self-supervised method for partial point set registration. While recent proposed learning-based methods have achieved impressive registration performance on the full shape observations, these methods mostly suffer from performance degradation when dealing with partial shapes. To bridge the performance gaps between partial point set registration with full point set registration, we proposed to incorporate a shape completion network to benefit the registration process. To achieve this, we design a latent code for each pair of shapes, which can be regarded as a geometric encoding of the target shape. By doing so, our model does need an explicit feature embedding network to learn the feature encodings. More importantly, both our shape completion network and the point set registration network take the shared latent codes as input, which are optimized along with the parameters of two decoder networks in the training process. Therefore, the point set registration process can thus benefit from the joint optimization process of latent codes, which are enforced to represent the information of full shape instead of partial ones. In the inference stage, we fix the network parameter and optimize the latent codes to get the optimal shape completion and registration results. Our proposed method is pure unsupervised and does not need any ground truth supervision. Experiments on the ModelNet40 dataset demonstrate the effectiveness of our model for partial point set registration.

Hao Huang, Jianchun Chen, Xiang Li et al.

Estimating dense correspondences between images is a long-standing image under-standing task. Recent works introduce convolutional neural networks (CNNs) to extract high-level feature maps and find correspondences through feature matching. However,high-level feature maps are in low spatial resolution and therefore insufficient to provide accurate and fine-grained features to distinguish intra-class variations for correspondence matching. To address this problem, we generate robust features by dynamically selecting features at different scales. To resolve two critical issues in feature selection,i.e.,how many and which scales of features to be selected, we frame the feature selection process as a sequential Markov decision-making process (MDP) and introduce an optimal selection strategy using reinforcement learning (RL). We define an RL environment for image matching in which each individual action either requires new features or terminates the selection episode by referring a matching score. Deep neural networks are incorporated into our method and trained for decision making. Experimental results show that our method achieves comparable/superior performance with state-of-the-art methods on three benchmarks, demonstrating the effectiveness of our feature selection strategy.

Lingjing Wang, Xiang Li, Yi Fang

In this paper, we propose a novel method named GP-Aligner to deal with the problem of non-rigid groupwise point set registration. Compared to previous non-learning approaches, our proposed method gains competitive advantages by leveraging the power of deep neural networks to effectively and efficiently learn to align a large number of highly deformed 3D shapes with superior performance. Unlike most learning-based methods that use an explicit feature encoding network to extract the per-shape features and their correlations, our model leverages a model-free learnable latent descriptor to characterize the group relationship. More specifically, for a given group we first define an optimizable Group Latent Descriptor (GLD) to characterize the gruopwise relationship among a group of point sets. Each GLD is randomly initialized from a Gaussian distribution and then concatenated with the coordinates of each point of the associated point sets in the group. A neural network-based decoder is further constructed to predict the coherent drifts as the desired transformation from input groups of shapes to aligned groups of shapes. During the optimization process, GP-Aligner jointly updates all GLDs and weight parameters of the decoder network towards the minimization of an unsupervised groupwise alignment loss. After optimization, for each group our model coherently drives each point set towards a middle, common position (shape) without specifying one as the target. GP-Aligner does not require large-scale training data for network training and it can directly align groups of point sets in a one-stage optimization process. GP-Aligner shows both accuracy and computational efficiency improvement in comparison with state-of-the-art methods for groupwise point set registration. Moreover, GP-Aligner is shown great efficiency in aligning a large number of groups of real-world 3D shapes.

Shuaihang Yuan, Xiang Li, Anthony Tzes et al.

In this paper, we deal with the problem to predict the future 3D motions of 3D object scans from previous two consecutive frames. Previous methods mostly focus on sparse motion prediction in the form of skeletons. While in this paper we focus on predicting dense 3D motions in the from of 3D point clouds. To approach this problem, we propose a self-supervised approach that leverages the power of the deep neural network to learn a continuous flow function of 3D point clouds that can predict temporally consistent future motions and naturally bring out the correspondences among consecutive point clouds at the same time. More specifically, in our approach, to eliminate the unsolved and challenging process of defining a discrete point convolution on 3D point cloud sequences to encode spatial and temporal information, we introduce a learnable latent code to represent the temporal-aware shape descriptor which is optimized during model training. Moreover, a temporally consistent motion Morpher is proposed to learn a continuous flow field which deforms a 3D scan from the current frame to the next frame. We perform extensive experiments on D-FAUST, SCAPE and TOSCA benchmark data sets and the results demonstrate that our approach is capable of handling temporally inconsistent input and produces consistent future 3D motion while requiring no ground truth supervision.

Shuaihang Yuan, Xiang Li, Yi Fang

This paper deals with the problem of 3D tracking, i.e., to find dense correspondences in a sequence of time-varying 3D shapes. Despite deep learning approaches have achieved promising performance for pairwise dense 3D shapes matching, it is a great challenge to generalize those approaches for the tracking of 3D time-varying geometries. In this paper, we aim at handling the problem of 3D tracking, which provides the tracking of the consecutive frames of 3D shapes. We propose a novel unsupervised 3D shape registration framework named DeepTracking-Net, which uses the deep neural networks (DNNs) as auxiliary functions to produce spatially and temporally continuous displacement fields for 3D tracking of objects in a temporal order. Our key novelty is that we present a novel temporal-aware correspondence descriptor (TCD) that captures spatio-temporal essence from consecutive 3D point cloud frames. Specifically, our DeepTracking-Net starts with optimizing a randomly initialized latent TCD. The TCD is then decoded to regress a continuous flow (i.e. a displacement vector field) which assigns a motion vector to every point of time-varying 3D shapes. Our DeepTracking-Net jointly optimizes TCDs and DNNs' weights towards the minimization of an unsupervised alignment loss. Experiments on both simulated and real data sets demonstrate that our unsupervised DeepTracking-Net outperforms the current supervised state-of-the-art method. In addition, we prepare a new synthetic 3D data, named SynMotions, to the 3D tracking and recognition community.

Lingjing Wang, Yi Shi, Xiang Li et al.

Global registration of point clouds aims to find an optimal alignment of a sequence of 2D or 3D point sets. In this paper, we present a novel method that takes advantage of current deep learning techniques for unsupervised learning of global registration from a temporal sequence of point clouds. Our key novelty is that we introduce a deep Spatio-Temporal REPresentation (STREP) feature, which describes the geometric essence of both temporal and spatial relationship of the sequence of point clouds acquired with sensors in an unknown environment. In contrast to the previous practice that treats each time step (pair-wise registration) individually, our unsupervised model starts with optimizing a sequence of latent STREP feature, which is then decoded to a temporally and spatially continuous sequence of geometric transformations to globally align multiple point clouds. We have evaluated our proposed approach over both simulated 2D and real 3D datasets and the experimental results demonstrate that our method can beat other techniques by taking into account the temporal information in deep feature learning.

Jingyu Deng, Xiang Li, Yi Fang

In this paper, we deal with the problem of object detection on remote sensing images. Previous methods have developed numerous deep CNN-based methods for object detection on remote sensing images and the report remarkable achievements in detection performance and efficiency. However, current CNN-based methods mostly require a large number of annotated samples to train deep neural networks and tend to have limited generalization abilities for unseen object categories. In this paper, we introduce a few-shot learning-based method for object detection on remote sensing images where only a few annotated samples are provided for the unseen object categories. More specifically, our model contains three main components: a meta feature extractor that learns to extract feature representations from input images, a reweighting module that learn to adaptively assign different weights for each feature representation from the support images, and a bounding box prediction module that carries out object detection on the reweighted feature maps. We build our few-shot object detection model upon YOLOv3 architecture and develop a multi-scale object detection framework. Experiments on two benchmark datasets demonstrate that with only a few annotated samples our model can still achieve a satisfying detection performance on remote sensing images and the performance of our model is significantly better than the well-established baseline models.

Lingjing Wang, Xiang Li, Yi Fang

Point cloud registration is the process of aligning a pair of point sets via searching for a geometric transformation. Recent works leverage the power of deep learning for registering a pair of point sets. However, unfortunately, deep learning models often require a large number of ground truth labels for training. Moreover, for a pair of source and target point sets, existing deep learning mechanisms require explicitly designed encoders to extract both deep spatial features from unstructured point clouds and their spatial correlation representation, which is further fed to a decoder to regress the desired geometric transformation for point set alignment. To further enhance deep learning models for point set registration, this paper proposes Deep-3DAligner, a novel unsupervised registration framework based on a newly introduced deep Spatial Correlation Representation (SCR) feature. The SCR feature describes the geometric essence of the spatial correlation between source and target point sets in an encoding-free manner. More specifically, our method starts with optimizing a randomly initialized latent SCR feature, which is then decoded to a geometric transformation (i.e., rotation and translation) to align source and target point sets. Our Deep-3DAligner jointly updates the SCR feature and weights of the transformation decoder towards the minimization of an unsupervised alignment loss. We conducted experiments on the ModelNet40 datasets to validate the performance of our unsupervised Deep-3DAligner for point set registration. The results demonstrated that, even without ground truth and any assumption of a direct correspondence between source and target point sets for training, our proposed approach achieved comparative performance compared to most recent supervised state-of-the-art approaches.

Xiang Li, Lingjing Wang, Yi Fang

Recent studies have shown the benefits of using additional elevation data (e.g., DSM) for enhancing the performance of the semantic segmentation of aerial images. However, previous methods mostly adopt 3D elevation information as additional inputs. While in many real-world applications, one does not have the corresponding DSM information at hand and the spatial resolution of acquired DSM images usually do not match the aerial images. To alleviate this data constraint and also take advantage of 3D elevation information, in this paper, we introduce a geometry-aware segmentation model that achieves accurate semantic labeling of aerial images via joint height estimation. Instead of using a single-stream encoder-decoder network for semantic labeling, we design a separate decoder branch to predict the height map and use the DSM images as side supervision to train this newly designed decoder branch. In this way, our model does not require DSM as model input and still benefits from the helpful 3D geometric information during training. Moreover, we develop a new geometry-aware convolution module that fuses the 3D geometric features from the height decoder branch and the 2D contextual features from the semantic segmentation branch. The fused feature embeddings can produce geometry-aware segmentation maps with enhanced performance. Our model is trained with DSM images as side supervision, while in the inference stage, it does not require DSM data and directly predicts the semantic labels in an end-to-end fashion. Experiments on ISPRS Vaihingen and Potsdam datasets demonstrate the effectiveness of the proposed method for the semantic segmentation of aerial images. The proposed model achieves remarkable performance on both datasets without using any hand-crafted features or post-processing.

Xiang Li, Mingyang Wang, Yi Fang

Understanding the 3D geometric structure of the Earth's surface has been an active research topic in photogrammetry and remote sensing community for decades, serving as an essential building block for various applications such as 3D digital city modeling, change detection, and city management. Previous researches have extensively studied the problem of height estimation from aerial images based on stereo or multi-view image matching. These methods require two or more images from different perspectives to reconstruct 3D coordinates with camera information provided. In this paper, we deal with the ambiguous and unsolved problem of height estimation from a single aerial image. Driven by the great success of deep learning, especially deep convolution neural networks (CNNs), some researches have proposed to estimate height information from a single aerial image by training a deep CNN model with large-scale annotated datasets. These methods treat height estimation as a regression problem and directly use an encoder-decoder network to regress the height values. In this paper, we proposed to divide height values into spacing-increasing intervals and transform the regression problem into an ordinal regression problem, using an ordinal loss for network training. To enable multi-scale feature extraction, we further incorporate an Atrous Spatial Pyramid Pooling (ASPP) module to extract features from multiple dilated convolution layers. After that, a post-processing technique is designed to transform the predicted height map of each patch into a seamless height map. Finally, we conduct extensive experiments on ISPRS Vaihingen and Potsdam datasets. Experimental results demonstrate significantly better performance of our method compared to the state-of-the-art methods.

Xiang Li, Mingyang Wang, Congcong Wen et al.

To better address challenging issues of the irregularity and inhomogeneity inherently present in 3D point clouds, researchers have been shifting their focus from the design of hand-craft point feature towards the learning of 3D point signatures using deep neural networks for 3D point cloud classification. Recent proposed deep learning based point cloud classification methods either apply 2D CNN on projected feature images or apply 1D convolutional layers directly on raw point sets. These methods cannot adequately recognize fine-grained local structures caused by the uneven density distribution of the point cloud data. In this paper, to address this challenging issue, we introduced a density-aware convolution module which uses the point-wise density to re-weight the learnable weights of convolution kernels. The proposed convolution module is able to fully approximate the 3D continuous convolution on unevenly distributed 3D point sets. Based on this convolution module, we further developed a multi-scale fully convolutional neural network with downsampling and upsampling blocks to enable hierarchical point feature learning. In addition, to regularize the global semantic context, we implemented a context encoding module to predict a global context encoding and formulated a context encoding regularizer to enforce the predicted context encoding to be aligned with the ground truth one. The overall network can be trained in an end-to-end fashion with the raw 3D coordinates as well as the height above ground as inputs. Experiments on the International Society for Photogrammetry and Remote Sensing (ISPRS) 3D labeling benchmark demonstrated the superiority of the proposed method for point cloud classification. Our model achieved a new state-of-the-art performance with an average F1 score of 71.2% and improved the performance by a large margin on several categories.

Jiaxin Chen, Yi Fang

Due to the large cross-modality discrepancy between 2D sketches and 3D shapes, retrieving 3D shapes by sketches is a significantly challenging task. To address this problem, we propose a novel framework to learn a discriminative deep cross-modality adaptation model in this paper. Specifically, we first separately adopt two metric networks, following two deep convolutional neural networks (CNNs), to learn modality-specific discriminative features based on an importance-aware metric learning method. Subsequently, we explicitly introduce a cross-modality transformation network to compensate for the divergence between two modalities, which can transfer features of 2D sketches to the feature space of 3D shapes. We develop an adversarial learning based method to train the transformation model, by simultaneously enhancing the holistic correlations between data distributions of two modalities, and mitigating the local semantic divergences through minimizing a cross-modality mean discrepancy term. Experimental results on the SHREC 2013 and SHREC 2014 datasets clearly show the superior retrieval performance of our proposed model, compared to the state-of-the-art approaches.

Lingjing Wang, Yi Fang

Recent advancements in deep learning opened new opportunities for learning a high-quality 3D model from a single 2D image given sufficient training on large-scale data sets. However, the significant imbalance between available amount of images and 3D models, and the limited availability of labeled 2D image data (i.e. manually annotated pairs between images and their corresponding 3D models), severely impacts the training of most supervised deep learning methods in practice. In this paper, driven by a novel design of adversarial networks, we have developed an unsupervised learning paradigm to reconstruct 3D models from a single 2D image, which is free of manually annotated pairwise input image and its associated 3D model. Particularly, the paradigm begins with training an adaption network via autoencoder with adversarial loss, which embeds unpaired 2D synthesized image domain with real world image domain to a shared latent vector space. Then, we jointly train a 3D deconvolutional network to transform the latent vector space to the 3D object space together with the embedding process. Our experiments verify our network's robust and superior performance in handling 3D volumetric object generation from a single 2D image.