Deep Convolutional Neural Networks (CNNs) have long been the architecture of choice for computer vision tasks. Recently, Transformer-based architectures like Vision Transformer (ViT) have matched or even surpassed ResNets for image classification. However, details of the Transformer architecture such as the use of non-overlapping patches lead one to wonder whether these networks are as robust. In this paper, we perform an extensive study of a variety of different measures of robustness of ViT models and compare the findings to ResNet baselines. We investigate robustness to input perturbations as well as robustness to model perturbations. We find that when pre-trained with a sufficient amount of data, ViT models are at least as robust as the ResNet counterparts on a broad range of perturbations. We also find that Transformers are robust to the removal of almost any single layer, and that while activations from later layers are highly correlated with each other, they nevertheless play an important role in classification.
论文检索
输入标题、作者或关键词，从 1,612 篇学术成果中精准定位
Deep generative models of 3D shapes have received a great deal of research interest. Yet, almost all of them generate discrete shape representations, such as voxels, point clouds, and polygon meshes. We present the first 3D generative model for a drastically different shape representation --- describing a shape as a sequence of computer-aided design (CAD) operations. Unlike meshes and point clouds, CAD models encode the user creation process of 3D shapes, widely used in numerous industrial and engineering design tasks. However, the sequential and irregular structure of CAD operations poses significant challenges for existing 3D generative models. Drawing an analogy between CAD operations and natural language, we propose a CAD generative network based on the Transformer. We demonstrate the performance of our model for both shape autoencoding and random shape generation. To train our network, we create a new CAD dataset consisting of 178,238 models and their CAD construction sequences. We have made this dataset publicly available to promote future research on this topic.
We present a method for reconstructing the 3D shape of underwater environments from a single, stationary camera placed above the water. We propose a novel differentiable framework, which, to our knowledge, is the first single-camera solution that is capable of simultaneously retrieving the structure of dynamic water surfaces and static underwater scene geometry in the wild. This framework integrates ray casting of Snell's law at the refractive interface, multi-view triangulation and specially designed loss functions. Our method is calibration-free, and thus it is easy to collect data outdoors in uncontrolled environments. Experimental results show that our method is able to realize robust and quality reconstructions on a variety of scenes, both in a laboratory environment and in the wild, and even in a salt water environment. We believe the method is promising for applications in surveying and environmental monitoring.
We introduce the first Neural Architecture Search (NAS) method to find a better transformer architecture for image recognition. Recently, transformers without CNN-based backbones are found to achieve impressive performance for image recognition. However, the transformer is designed for NLP tasks and thus could be sub-optimal when directly used for image recognition. In order to improve the visual representation ability for transformers, we propose a new search space and searching algorithm. Specifically, we introduce a locality module that models the local correlations in images explicitly with fewer computational cost. With the locality module, our search space is defined to let the search algorithm freely trade off between global and local information as well as optimizing the low-level design choice in each module. To tackle the problem caused by huge search space, a hierarchical neural architecture search method is proposed to search the optimal vision transformer from two levels separately with the evolutionary algorithm. Extensive experiments on the ImageNet dataset demonstrate that our method can find more discriminative and efficient transformer variants than the ResNet family (e.g., ResNet101) and the baseline ViT for image classification.
Marker-based optical motion capture (mocap) is the "gold standard" method for acquiring accurate 3D human motion in computer vision, medicine, and graphics. The raw output of these systems are noisy and incomplete 3D points or short tracklets of points. To be useful, one must associate these points with corresponding markers on the captured subject; i.e. "labelling". Given these labels, one can then "solve" for the 3D skeleton or body surface mesh. Commercial auto-labeling tools require a specific calibration procedure at capture time, which is not possible for archival data. Here we train a novel neural network called SOMA, which takes raw mocap point clouds with varying numbers of points, labels them at scale without any calibration data, independent of the capture technology, and requiring only minimal human intervention. Our key insight is that, while labeling point clouds is highly ambiguous, the 3D body provides strong constraints on the solution that can be exploited by a learning-based method. To enable learning, we generate massive training sets of simulated noisy and ground truth mocap markers animated by 3D bodies from AMASS. SOMA exploits an architecture with stacked self-attention elements to learn the spatial structure of the 3D body and an optimal transport layer to constrain the assignment (labeling) problem while rejecting outliers. We extensively evaluate SOMA both quantitatively and qualitatively. SOMA is more accurate and robust than existing state of the art research methods and can be applied where commercial systems cannot. We automatically label over 8 hours of archival mocap data across 4 different datasets captured using various technologies and output SMPL-X body models. The model and data is released for research purposes at https://soma.is.tue.mpg.de/.
In recent years, deep learning moves video-based Continuous Sign Language Recognition (CSLR) significantly forward. Currently, a typical network combination for CSLR includes a visual module, which focuses on spatial and short-temporal information, followed by a contextual module, which focuses on long-temporal information, and the Connectionist Temporal Classification (CTC) loss is adopted to train the network. However, due to the limitation of chain rules in back-propagation, the visual module is hard to adjust for seeking optimized visual features. As a result, it enforces that the contextual module focuses on contextual information optimization only rather than balancing efficient visual and contextual information. In this paper, we propose a Self-Mutual Knowledge Distillation (SMKD) method, which enforces the visual and contextual modules to focus on short-term and long-term information and enhances the discriminative power of both modules simultaneously. Specifically, the visual and contextual modules share the weights of their corresponding classifiers, and train with CTC loss simultaneously. Moreover, the spike phenomenon widely exists with CTC loss. Although it can help us choose a few of the key frames of a gloss, it does drop other frames in a gloss and makes the visual feature saturation in the early stage. A gloss segmentation is developed to relieve the spike phenomenon and decrease saturation in the visual module. We conduct experiments on two CSLR benchmarks: PHOENIX14 and PHOENIX14-T. Experimental results demonstrate the effectiveness of the SMKD.
Weakly supervised object localization (WSOL) aims to localize objects with only image-level labels, which has better scalability and practicability than fully supervised methods in the actual deployment. However, with only image-level labels, learning object classification models tends to activate object parts and ignore the whole object, while expanding object parts into the whole object may deteriorate classification performance. To alleviate this problem, we propose foreground activation maps (FAM), whose aim is to optimize object localization and classification jointly via an object-aware attention module and a part-aware attention module in a unified model, where the two tasks can complement and enhance each other. To the best of our knowledge, this is the first work that can achieve remarkable performance for both tasks by optimizing them jointly via FAM for WSOL. Besides, the designed two modules can effectively highlight foreground objects for localization and discover discriminative parts for classification. Extensive experiments with four backbones on two standard benchmarks demonstrate that our FAM performs favorably against state-of-the-art WSOL methods.
Batch Normalization Increases Adversarial Vulnerability and Decreases Adversarial Transferability: A Non-Robust Feature Perspective
PDF ↗Batch normalization (BN) has been widely used in modern deep neural networks (DNNs) due to improved convergence. BN is observed to increase the model accuracy while at the cost of adversarial robustness. There is an increasing interest in the ML community to understand the impact of BN on DNNs, especially related to the model robustness. This work attempts to understand the impact of BN on DNNs from a non-robust feature perspective. Straightforwardly, the improved accuracy can be attributed to the better utilization of useful features. It remains unclear whether BN mainly favors learning robust features (RFs) or non-robust features (NRFs). Our work presents empirical evidence that supports that BN shifts a model towards being more dependent on NRFs. To facilitate the analysis of such a feature robustness shift, we propose a framework for disentangling robust usefulness into robustness and usefulness. Extensive analysis under the proposed framework yields valuable insight on the DNN behavior regarding robustness, e.g. DNNs first mainly learn RFs and then NRFs. The insight that RFs transfer better than NRFs, further inspires simple techniques to strengthen transfer-based black-box attacks.
Extracting robust feature representation is one of the key challenges in object re-identification (ReID). Although convolution neural network (CNN)-based methods have achieved great success, they only process one local neighborhood at a time and suffer from information loss on details caused by convolution and downsampling operators pooling and strided convolution).To overcome these limitations, we propose a pure transformer-based object ReID framework named TransReID. Specifically, we first encode an image as a sequence of patches and build a transformer-based strong baseline with a few critical improvements, which achieves competitive results on several ReID benchmarks with CNN-based methods. To further enhance the robust feature learning in the context of transformers, two novel modules are carefully designed. (i) The jigsaw patch module (JPM) is proposed to rearrange the patch embeddings via shift and patch shuffle operations which generates robust features with improved discrimination ability and more diversified coverage. (ii) The side information embeddings (SIE) is introduced to mitigate feature bias towards camera/view variations by plugging in learnable embeddings to incorporate these non-visual clues. To the best of our knowledge, this is the first work to adopt a pure transformer for ReID research. Experimental results of TransReID are superior promising, which achieve state-of-the-art performance on both person and vehicle ReID benchmarks. Code is available at https://github.com/heshuting555/TransReID
Recent advances in deep generative models have led to immense progress in 3D shape synthesis. While existing models are able to synthesize shapes represented as voxels, point-clouds, or implicit functions, these methods only indirectly enforce the plausibility of the final 3D shape surface. Here we present a 3D shape synthesis framework (SurfGen) that directly applies adversarial training to the object surface. Our approach uses a differentiable spherical projection layer to capture and represent the explicit zero isosurface of an implicit 3D generator as functions defined on the unit sphere. By processing the spherical representation of 3D object surfaces with a spherical CNN in an adversarial setting, our generator can better learn the statistics of natural shape surfaces. We evaluate our model on large-scale shape datasets, and demonstrate that the end-to-end trained model is capable of generating high fidelity 3D shapes with diverse topology
Semi-supervised approaches for crowd counting attract attention, as the fully supervised paradigm is expensive and laborious due to its request for a large number of images of dense crowd scenarios and their annotations. This paper proposes a spatial uncertainty-aware semi-supervised approach via regularized surrogate task (binary segmentation) for crowd counting problems. Different from existing semi-supervised learning-based crowd counting methods, to exploit the unlabeled data, our proposed spatial uncertainty-aware teacher-student framework focuses on high confident regions' information while addressing the noisy supervision from the unlabeled data in an end-to-end manner. Specifically, we estimate the spatial uncertainty maps from the teacher model's surrogate task to guide the feature learning of the main task (density regression) and the surrogate task of the student model at the same time. Besides, we introduce a simple yet effective differential transformation layer to enforce the inherent spatial consistency regularization between the main task and the surrogate task in the student model, which helps the surrogate task to yield more reliable predictions and generates high-quality uncertainty maps. Thus, our model can also address the task-level perturbation problems that occur spatial inconsistency between the primary and surrogate tasks in the student model. Experimental results on four challenging crowd counting datasets demonstrate that our method achieves superior performance to the state-of-the-art semi-supervised methods. Code is available at : https://github.com/smallmax00/SUA_crowd_counting
High-quality point clouds have practical significance for point-based rendering, semantic understanding, and surface reconstruction. Upsampling sparse, noisy and non-uniform point clouds for a denser and more regular approximation of target objects is a desirable but challenging task. Most existing methods duplicate point features for upsampling, constraining the upsampling scales at a fixed rate. In this work, the arbitrary point clouds upsampling rates are achieved via edge-vector based affine combinations, and a novel design of Edge-Vector based Approximation for Flexible-scale Point clouds Upsampling (PU-EVA) is proposed. The edge-vector based approximation encodes neighboring connectivity via affine combinations based on edge vectors, and restricts the approximation error within a second-order term of Taylor's Expansion. Moreover, the EVA upsampling decouples the upsampling scales with network architecture, achieving the arbitrary upsampling rates in one-time training. Qualitative and quantitative evaluations demonstrate that the proposed PU-EVA outperforms the state-of-the-arts in terms of proximity-to-surface, distribution uniformity, and geometric details preservation.
We present AIST++, a new multi-modal dataset of 3D dance motion and music, along with FACT, a Full-Attention Cross-modal Transformer network for generating 3D dance motion conditioned on music. The proposed AIST++ dataset contains 1.1M frames of 3D dance motion in 1408 sequences, covering 10 dance genres with multi-view videos with known camera poses---the largest dataset of this kind to our knowledge. We show that naively applying sequence models such as transformers to this dataset for the task of music conditioned 3D motion generation does not produce satisfactory 3D motion that is well correlated with the input music. We overcome these shortcomings by introducing key changes in its architecture design and supervision: FACT model involves a deep cross-modal transformer block with full-attention that is trained to predict N future motions. We empirically show that these changes are key factors in generating long sequences of realistic dance motion that are well-attuned to the input music. We conduct extensive experiments on AIST++ with user studies, where our method outperforms recent state-of-the-art methods both qualitatively and quantitatively. The code and the dataset can be found at: https://google.github.io/aichoreographer.
We propose DeepMultiCap, a novel method for multi-person performance capture using sparse multi-view cameras. Our method can capture time varying surface details without the need of using pre-scanned template models. To tackle with the serious occlusion challenge for close interacting scenes, we combine a recently proposed pixel-aligned implicit function with parametric model for robust reconstruction of the invisible surface areas. An effective attention-aware module is designed to obtain the fine-grained geometry details from multi-view images, where high-fidelity results can be generated. In addition to the spatial attention method, for video inputs, we further propose a novel temporal fusion method to alleviate the noise and temporal inconsistencies for moving character reconstruction. For quantitative evaluation, we contribute a high quality multi-person dataset, MultiHuman, which consists of 150 static scenes with different levels of occlusions and ground truth 3D human models. Experimental results demonstrate the state-of-the-art performance of our method and the well generalization to real multiview video data, which outperforms the prior works by a large margin.
In this paper, we introduce a new framework for unsupervised deep homography estimation. Our contributions are 3 folds. First, unlike previous methods that regress 4 offsets for a homography, we propose a homography flow representation, which can be estimated by a weighted sum of 8 pre-defined homography flow bases. Second, considering a homography contains 8 Degree-of-Freedoms (DOFs) that is much less than the rank of the network features, we propose a Low Rank Representation (LRR) block that reduces the feature rank, so that features corresponding to the dominant motions are retained while others are rejected. Last, we propose a Feature Identity Loss (FIL) to enforce the learned image feature warp-equivariant, meaning that the result should be identical if the order of warp operation and feature extraction is swapped. With this constraint, the unsupervised optimization is achieved more effectively and more stable features are learned. Extensive experiments are conducted to demonstrate the effectiveness of all the newly proposed components, and results show that our approach outperforms the state-of-the-art on the homography benchmark datasets both qualitatively and quantitatively. Code is available at https://github.com/megvii-research/BasesHomo
When a camera is pointed at a strong light source, the resulting photograph may contain lens flare artifacts. Flares appear in a wide variety of patterns (halos, streaks, color bleeding, haze, etc.) and this diversity in appearance makes flare removal challenging. Existing analytical solutions make strong assumptions about the artifact's geometry or brightness, and therefore only work well on a small subset of flares. Machine learning techniques have shown success in removing other types of artifacts, like reflections, but have not been widely applied to flare removal due to the lack of training data. To solve this problem, we explicitly model the optical causes of flare either empirically or using wave optics, and generate semi-synthetic pairs of flare-corrupted and clean images. This enables us to train neural networks to remove lens flare for the first time. Experiments show our data synthesis approach is critical for accurate flare removal, and that models trained with our technique generalize well to real lens flares across different scenes, lighting conditions, and cameras.
Deep metric learning has been effectively used to learn distance metrics for different visual tasks like image retrieval, clustering, etc. In order to aid the training process, existing methods either use a hard mining strategy to extract the most informative samples or seek to generate hard synthetics using an additional network. Such approaches face different challenges and can lead to biased embeddings in the former case, and (i) harder optimization (ii) slower training speed (iii) higher model complexity in the latter case. In order to overcome these challenges, we propose a novel approach that looks for optimal hard negatives (LoOp) in the embedding space, taking full advantage of each tuple by calculating the minimum distance between a pair of positives and a pair of negatives. Unlike mining-based methods, our approach considers the entire space between pairs of embeddings to calculate the optimal hard negatives. Extensive experiments combining our approach and representative metric learning losses reveal a significant boost in performance on three benchmark datasets.
Gaze following, i.e., detecting the gaze target of a human subject, in 2D images has become an active topic in computer vision. However, it usually suffers from the out of frame issue due to the limited field-of-view (FoV) of 2D images. In this paper, we introduce a novel task, gaze following in 360-degree images which provide an omnidirectional FoV and can alleviate the out of frame issue. We collect the first dataset, "GazeFollow360", for this task, containing around 10,000 360-degree images with complex gaze behaviors under various scenes. Existing 2D gaze following methods suffer from performance degradation in 360-degree images since they may use the assumption that a gaze target is in the 2D gaze sight line. However, this assumption is no longer true for long-distance gaze behaviors in 360-degree images, due to the distortion brought by sphere-to-plane projection. To address this challenge, we propose a 3D sight line guided dual-pathway framework, to detect the gaze target within a local region (here) and from a distant region (there), parallelly. Specifically, the local region is obtained as a 2D cone-shaped field along the 2D projection of the sight line starting at the human subject's head position, and the distant region is obtained by searching along the sight line in 3D sphere space. Finally, the location of the gaze target is determined by fusing the estimations from both the local region and the distant region. Experimental results show that our method achieves significant improvements over previous 2D gaze following methods on our GazeFollow360 dataset.
3D multi-object tracking in LiDAR point clouds is a key ingredient for self-driving vehicles. Existing methods are predominantly based on the tracking-by-detection pipeline and inevitably require a heuristic matching step for the detection association. In this paper, we present SimTrack to simplify the hand-crafted tracking paradigm by proposing an end-to-end trainable model for joint detection and tracking from raw point clouds. Our key design is to predict the first-appear location of each object in a given snippet to get the tracking identity and then update the location based on motion estimation. In the inference, the heuristic matching step can be completely waived by a simple read-off operation. SimTrack integrates the tracked object association, newborn object detection, and dead track killing in a single unified model. We conduct extensive evaluations on two large-scale datasets: nuScenes and Waymo Open Dataset. Experimental results reveal that our simple approach compares favorably with the state-of-the-art methods while ruling out the heuristic matching rules.
Superpoints are formed by grouping similar points with local geometric structures, which can effectively reduce the number of primitives of point clouds for subsequent point cloud processing. Existing superpoint methods mainly focus on employing clustering or graph partition to generate superpoints with handcrafted or learned features. Nonetheless, these methods cannot learn superpoints of point clouds with an end-to-end network. In this paper, we develop a new deep iterative clustering network to directly generate superpoints from irregular 3D point clouds in an end-to-end manner. Specifically, in our clustering network, we first jointly learn a soft point-superpoint association map from the coordinate and feature spaces of point clouds, where each point is assigned to the superpoint with a learned weight. Furthermore, we then iteratively update the association map and superpoint centers so that we can more accurately group the points into the corresponding superpoints with locally similar geometric structures. Finally, by predicting the pseudo labels of the superpoint centers, we formulate a label consistency loss on the points and superpoint centers to train the network. Extensive experiments on various datasets indicate that our method not only achieves the state-of-the-art on superpoint generation but also improves the performance of point cloud semantic segmentation. Code is available at https://github.com/fpthink/SPNet.