CVPR 2021 论文列表

Spatial-temporal, channel-wise, and motion patterns are three complementary and crucial types of information for video action recognition. Conventional 2D CNNs are computationally cheap but cannot catch temporal relationships; 3D CNNs can achieve good performance but are computationally intensive. In this work, we tackle this dilemma by designing a generic and effective module that can be embedded into 2D CNNs. To this end, we propose a spAtio-temporal, Channel and moTion excitatION (ACTION) module consisting of three paths: Spatio-Temporal Excitation (STE) path, Channel Excitation (CE) path, and Motion Excitation (ME) path. The STE path employs one channel 3D convolution to characterize spatio-temporal representation. The CE path adaptively recalibrates channel-wise feature responses by explicitly modeling interdependencies between channels in terms of the temporal aspect. The ME path calculates feature-level temporal differences, which is then utilized to excite motion-sensitive channels. We equip 2D CNNs with the proposed ACTION module to form a simple yet effective ACTION-Net with very limited extra computational cost. ACTION-Net is demonstrated by consistently outperforming 2D CNN counterparts on three backbones (i.e., ResNet-50, MobileNet V2 and BNInception) employing three datasets (i.e., Something-Something V2, Jester, and EgoGesture). Code is provided at https://github.com/V-Sense/ACTION-Net.

Causal induction, i.e., identifying unobservable mechanisms that lead to the observable relations among variables, has played a pivotal role in modern scientific discovery, especially in scenarios with only sparse and limited data. Humans, even young toddlers, can induce causal relationships surprisingly well in various settings despite its notorious difficulty. However, in contrast to the commonplace trait of human cognition is the lack of a diagnostic benchmark to measure causal induction for modern Artificial Intelligence (AI) systems. Therefore, in this work, we introduce the Abstract Causal REasoning (ACRE) dataset for systematic evaluation of current vision systems in causal induction. Motivated by the stream of research on causal discovery in Blicket experiments, we query a visual reasoning system with the following four types of questions in either an independent scenario or an interventional scenario: direct, indirect, screening-off, and backward-blocking, intentionally going beyond the simple strategy of inducing causal relationships by covariation. By analyzing visual reasoning architectures on this testbed, we notice that pure neural models tend towards an associative strategy under their chance-level performance, whereas neuro-symbolic combinations struggle in backward-blocking reasoning. These deficiencies call for future research in models with a more comprehensive capability of causal induction.

Semantic segmentation models gain robustness against poor lighting conditions by virtue of complementary information from visible (RGB) and thermal images. Despite its importance, most existing RGB-T semantic segmentation models perform primitive fusion strategies, such as concatenation, element-wise summation and weighted summation, to fuse features from different modalities. These strategies, unfortunately, overlook the modality differences due to different imaging mechanisms, so that they suffer from the reduced discriminability of the fused features. To address such an issue, we propose, for the first time, the strategy of bridging-then-fusing, where the innovation lies in a novel Adaptive-weighted Bi-directional Modality Difference Reduction Network (ABMDRNet). Concretely, a Modality Difference Reduction and Fusion (MDRF) subnetwork is designed, which first employs a bi-directional image-to-image translation based method to reduce the modality differences between RGB features and thermal features, and then adaptively selects those discriminative multi-modality features for RGB-T semantic segmentation in a channel-wise weighted fusion way. Furthermore, considering the importance of contextual information in semantic segmentation, a Multi-Scale Spatial Context (MSC) module and a Multi-Scale Channel Context (MCC) module are proposed to exploit the interactions among multi-scale contextual information of cross-modality features together with their long-range dependencies along spatial and channel dimensions, respectively. Comprehensive experiments on MFNet dataset demonstrate that our method achieves new state-of-the-art results.

Learning pyramidal feature representations is crucial for recognizing object instances at different scales. Feature Pyramid Network (FPN) is the classic architecture to build a feature pyramid with high-level semantics throughout. However, intrinsic defects in feature extraction and fusion inhibit FPN from further aggregating more discriminative features. In this work, we propose Attention Aggregation based Feature Pyramid Network (A^2-FPN), to improve multi-scale feature learning through attention-guided feature aggregation. In feature extraction, it extracts discriminative features by collecting-distributing multi-level global context features, and mitigates the semantic information loss due to drastically reduced channels. In feature fusion, it aggregates complementary information from adjacent features to generate location-wise reassembly kernels for content-aware sampling, and employs channel-wise reweighting to enhance the semantic consistency before element-wise addition. A^2-FPN shows consistent gains on different instance segmentation frameworks. By replacing FPN with A^2-FPN in Mask R-CNN, our model boosts the performance by 2.1% and 1.6% mask AP when using ResNet-50 and ResNet-101 as backbone, respectively. Moreover, A^2-FPN achieves an improvement of 2.0% and 1.4% mask AP when integrated into the strong baselines such as Cascade Mask R-CNN and Hybrid Task Cascade.

We address the problem of computing a textural loss based on the statistics extracted from the feature activations of a convolutional neural network optimized for object recognition (e.g. VGG-19). The underlying mathematical problem is the measure of the distance between two distributions in feature space. The Gram-matrix loss is the ubiquitous approximation for this problem but it is subject to several shortcomings. Our goal is to promote the Sliced Wasserstein Distance as a replacement for it. It is theoretically proven, practical, simple to implement, and achieves results that are visually superior for texture synthesis by optimization or training generative neural networks.

Automated radiographic report generation is a challenging task since it requires to generate paragraphs describing fine-grained visual differences of cases, especially for those between the diseased and the healthy. Existing image captioning methods commonly target at generic images, and lack mechanism to meet this requirement. To bridge this gap, in this paper, we propose a self-boosting framework that improves radiographic report generation based on the cooperation of the main task of report generation and anauxiliary task of image-text matching. The two tasks are built as the two branches of a network model and influence each other in a cooperative way. On one hand, the image-text matching branch helps to learn highly text-correlated visual features for the report generation branch to output high quality reports. One the other hand, the improved reports produced by the report generation branch provideadditional harder samples for the image-text matching task and enforce the latter to improve itself by learning better visual and text feature representations. This, in turn, helps improve the report generation branch again. These two branches are jointly trained to help improve each other iteratively and progressively, so that the whole model is self-boosted without requiring any external resources. Additionally, in the loss function, our model evaluates the quality of the generated reports not only on the word similarity as common approaches do (via minimizing a cross-entropy loss), but also on the feature similarity at high-level, while the latter is provided by the text-encoder of the image-text matching branch. Experimental results demonstrate the effectiveness of our method on two public datasets, showing its superior performance over other state-of-the-art medical report generation methods.

The presence of label noise often misleads the training of deep neural networks. Departing from the recent literature which largely assumes the label noise rate is only determined by the true label class, the errors in human-annotated labels are more likely to be dependent on the difficulty levels of tasks, resulting in settings with instance-dependent label noise. We first provide evidences that the heterogeneous instance-dependent label noise is effectively down-weighting the examples with higher noise rates in a non-uniform way and thus causes imbalances, rendering the strategy of directly applying methods for class-dependent label noise questionable. Built on a recent work peer loss [24], we then propose and study the potentials of a second-order approach that leverages the estimation of several covariance terms defined between the instance-dependent noise rates and the Bayes optimal label. We show that this set of second-order statistics successfully captures the induced imbalances. We further proceed to show that with the help of the estimated second-order statistics, we identify a new loss function whose expected risk of a classifier under instance-dependent label noise is equivalent to a new problem with only class-dependent label noise. This fact allows us to apply existing solutions to handle this better-studied setting. We provide an efficient procedure to estimate these second-order statistics without accessing either ground truth labels or prior knowledge of the noise rates. Experiments on CIFAR10 and CIFAR100 with synthetic instance-dependent label noise and Clothing1M with real-world human label noise verify our approach. Our implementation is available at https://github.com/UCSC-REAL/CAL.

We evaluate the effectiveness of semi-supervised learning (SSL) on a realistic benchmark where data exhibits considerable class imbalance and contains images from novel classes. Our benchmark consists of two fine-grained classification datasets obtained by sampling classes from the Aves and Fungi taxonomy. We find that recently proposed SSL methods provide significant benefits, and can effectively use out-of-class data to improve performance when deep networks are trained from scratch. Yet their performance pales in comparison to a transfer learning baseline, an alternative approach for learning from a few examples. Furthermore, in the transfer setting, while existing SSL methods provide improvements, the presence of out-of-class is often detrimental. In this setting, standard fine-tuning followed by distillation-based self-training is the most robust. Our work suggests that semi-supervised learning with experts on realistic datasets may require different strategies than those currently prevalent in the literature.

In this paper we study structure from motion problems for 1D radial cameras. Under this model the projection of a 3D point is a line in the image plane going through the principal point, which makes the model invariant to radial distortion and changes in focal length. It can therefore effectively be applied to uncalibrated image collections without the need for explicit estimation of camera intrinsics. We show that the reprojection errors of 1D radial cameras are examples of quasiconvex functions. This opens up the possibility to solve a general class of relevant reconstruction problems globally optimally using tools from convex optimization. In fact, our resulting algorithm is based on solving a series of LP problems. We perform an extensive experimental evaluation, on both synthetic and real data, showing that a whole class of multiview geometry problems across a range of different cameras models with varying and unknown intrinsic calibration can be reliably and accurately solved within the same framework.

Despite substantial progress in applying neural networks (NN) to a wide variety of areas, they still largely suffer from a lack of transparency and interpretability. While recent developments in explainable artificial intelligence attempt to bridge this gap (e.g., by visualizing the correlation between input pixels and final outputs), these approaches are limited to explaining low-level relationships, and crucially, do not provide insights on error correction. In this work, we propose a framework (VRX) to interpret classification NNs with intuitive structural visual concepts. Given a trained classification model, the proposed VRX extracts relevant class-specific visual concepts and organizes them using structural concept graphs (SCG) based on pairwise concept relationships. By means of knowledge distillation, we show VRX can take a step towards mimicking the reasoning process of NNs and provide logical, concept-level explanations for final model decisions. With extensive experiments, we empirically show VRX can meaningfully answer "why" and "why not" questions about the prediction, providing easy-to-understand insights about the reasoning process. We also show that these insights can potentially provide guidance on improving NN's performance.

Recent advances in OCR have shown that an end-to-end (E2E) training pipeline that includes both detection and recognition leads to the best results. However, many existing methods focus primarily on Latin-alphabet languages, often even only case-insensitive English characters. In this paper, we propose an E2E approach, Multiplexed Multilingual Mask TextSpotter, that performs script identification at the word level and handles different scripts with different recognition heads, all while maintaining a unified loss that simultaneously optimizes script identification and multiple recognition heads. Experiments show that our method outperforms single-head model with similar parameters in end-to-end recognition tasks, and achieves state-of-the-art results on MLT17 and MLT19 joint text detection and script identification benchmarks. We believe that our work is a step towards end-to-end trainable and scalable multilingual multi-purpose OCR system.

Specular highlight detection and removal are fundamental and challenging tasks. Although recent methods achieve promising results on the two tasks by supervised training on synthetic training data, they are typically solely designed for highlight detection or removal, and their performance usually deteriorates significantly on real-world images. In this paper, we present a novel network that aims to detect and remove highlights from natural images. To remove the domain gap between synthetic training samples and real test images, and support the investigation of learning-based approaches, we first introduce a dataset of 16K real images, each of which has the corresponding highlight detection and removal images. Using the presented dataset, we develop a multi-task network for joint highlight detection and removal, based on a new specular highlight image formation model. Experiments on the benchmark datasets and our new dataset show that our approach clearly outperforms the state-of-the-art methods for both highlight detection and removal.

We present a large-scale study on unsupervised spatiotemporal representation learning from videos. With a unified perspective on four recent image-based frameworks, we study a simple objective that can easily generalize all these methods to space-time. Our objective encourages temporally-persistent features in the same video, and in spite of its simplicity, it works surprisingly well across: (i) different unsupervised frameworks, (ii) pre-training datasets, (iii) downstream datasets, and (iv) backbone architectures. We draw a series of intriguing observations from this study, e.g., we discover that encouraging long-spanned persistency can be effective even if the timespan is 60 seconds. In addition to state-of-the-art results in multiple benchmarks, we report a few promising cases in which unsupervised pre-training can outperform its supervised counterpart. Code will be made available at https://github.com/facebookresearch/SlowFast.

Hyperbolic graph convolutional networks (GCNs) demonstrate powerful representation ability to model graphs with hierarchical structure. Existing hyperbolic GCNs resort to tangent spaces to realize graph convolution on hyperbolic manifolds, which is inferior because tangent space is only a local approximation of a manifold. In this paper, we propose a hyperbolic-to-hyperbolic graph convolutional network (H2H-GCN) that directly works on hyperbolic manifolds. Specifically, we developed a manifold-preserving graph convolution that consists of a hyperbolic feature transformation and a hyperbolic neighborhood aggregation. The hyperbolic feature transformation works as linear transformation on hyperbolic manifolds. It ensures the transformed node representations still lie on the hyperbolic manifold by imposing the orthogonal constraint on the transformation sub-matrix. The hyperbolic neighborhood aggregation updates each node representation via the Einstein midpoint. The H2H-GCN avoids the distortion caused by tangent space approximations and keeps the global hyperbolic structure. Extensive experiments show that the H2H-GCN achieves substantial improvements on the link prediction, node classification, and graph classification tasks.

Previous work shows that a better density map representation can improve the performance of crowd counting. In this paper, we investigate learning the density map representation through an unbalanced optimal transport problem, and propose a generalized loss function to learn density maps for crowd counting and localization. We prove that pixel-wise L2 loss and Bayesian loss are special cases and suboptimal solutions to our proposed loss function. A perspective-guided transport cost function is further proposed to better handle the perspective transformation in crowd images. Since the predicted density will be pushed toward annotation positions, the density map prediction will be sparse and can naturally be used for localization. Finally, the proposed loss outperforms other losses on four large-scale datasets for counting, and achieves the best localization performance on NWPU-Crowd and UCF-QNRF.

Learning pose invariant representation is a fundamental problem in shape analysis. Most existing deep learning algorithms for 3D shape analysis are not robust to rotations and are often trained on synthetic datasets consisting of pre-aligned shapes, yielding poor generalization to unseen poses. This observation motivates a growing interest in rotation invariant and equivariant methods. The field of rotation equivariant deep learning is developing in recent years thanks to a well established theory of Lie group representations and convolutions. A fundamental problem in equivariant deep learning is to design activation functions which are both informative and preserve equivariance. The recently introduced Tensor Field Network (TFN) framework provides a rotation equivariant network design for point cloud analysis. TFN features undergo a rotation in feature space given a rotation of the input pointcloud. TFN and similar designs consider nonlinearities which operate only over rotation invariant features such as the norm of equivariant features to preserve equivariance, making them unable to capture the directional information. In a recent work entitled "Gauge Equivariant Mesh CNNs: Anisotropic Convolutions on Geometric Graphs" Hann et al. interpret 2D rotation equivariant features as Fourier coefficients of functions on the circle. In this work we transpose the idea of Hann et al. to 3D by interpreting TFN features as spherical harmonics coefficients of functions on the sphere. We introduce a new equivariant nonlinearity and pooling for TFN. We show improvments over the original TFN design and other equivariant nonlinearities in classification and segmentation tasks. Furthermore our method is competitive with state of the art rotation invariant methods in some instances.

Modern deep neural networks suffer from performance degradation when evaluated on testing data under different distributions from training data. Domain generalization aims at tackling this problem by learning transferable knowledge from multiple source domains in order to generalize to unseen target domains. This paper introduces a novel Fourier-based perspective for domain generalization. The main assumption is that the Fourier phase information contains high-level semantics and is not easily affected by domain shifts. To force the model to capture phase information, we develop a novel Fourier-based data augmentation strategy called amplitude mix which linearly interpolates between the amplitude spectrums of two images. A dual-formed consistency loss called co-teacher regularization is further introduced between the predictions induced from original and augmented images. Extensive experiments on three benchmarks have demonstrated that the proposed method is able to achieve state-of-the-arts performance for domain generalization.

In this work, a robust and efficient dual iterative refinement (DIR) method is proposed for dense correspondence between two nearly isometric shapes. The key idea is to use dual information, such as spatial and spectral, or local and global features, in a complementary and effective way, and extract more accurate information from current iteration to use for the next iteration. In each DIR iteration, starting from current correspondence, a zoom-in process at each point is used to select well matched anchor pairs by a local mapping distortion criterion. These selected anchor pairs are then used to align spectral features (or other appropriate global features) whose dimension adaptively matches the capacity of the selected anchor pairs. Thanks to the effective combination of complementary information in a data-adaptive way, DIR is not only efficient but also robust to render accurate results within a few iterations. By choosing appropriate dual features, DIR has the flexibility to handle patch and partial matching as well. Extensive experiments on various data sets demonstrate the superiority of DIR over other state-of-the-art methods in terms of both accuracy and efficiency.

Fast and light-weight methods for animating 3D characters are desirable in various applications such as computer games. We present a learning-based approach to enhance skinning-based animations of 3D characters with vivid secondary motion effects. We represent each local patch of a character simulation mesh as a graph network where the edges implicitly encode the internal forces between the neighboring vertices. We then train a neural network that emulates the ordinary differential equations of the character dynamics, predicting new vertex positions from the current accelerations, velocities and positions. Being a local method, our network is independent of the mesh topology and generalizes to arbitrarily shaped 3D character meshes at test time. We further represent per-vertex constraints and material properties such as stiffness, enabling us to easily adjust the dynamics in different parts of the mesh. We evaluate our method on various character meshes and complex motion sequences. Our method can be over 30 times more efficient than ground-truth physically based simulation, and outperforms alternative solutions that provide fast approximations.