CVPR 2019 论文列表

Most state-of-the-art methods for action recognition consist of a two-stream architecture with 3D convolutions: an appearance stream for RGB frames and a motion stream for optical flow frames. Although combining flow with RGB improves the performance, the cost of computing accurate optical flow is high, and increases action recognition latency. This limits the usage of two-stream approaches in real-world applications requiring low latency. In this paper, we introduce two learning approaches to train a standard 3D CNN, operating on RGB frames, that mimics the motion stream, and as a result avoids flow computation at test time. First, by minimizing a feature-based loss compared to the Flow stream, we show that the network reproduces the motion stream with high fidelity. Second, to leverage both appearance and motion information effectively, we train with a linear combination of the feature-based loss and the standard cross-entropy loss for action recognition. We denote the stream trained using this combined loss as Motion-Augmented RGB Stream (MARS). As a single stream, MARS performs better than RGB or Flow alone, for instance with 72.7% accuracy on Kinetics compared to 72.0% and 65.6% with RGB and Flow streams respectively.

This research strives for natural language moment retrieval in long, untrimmed video streams. The problem is not trivial especially when a video contains multiple moments of interests and the language describes complex temporal dependencies, which often happens in real scenarios. We identify two crucial challenges: semantic misalignment and structural misalignment. However, existing approaches treat different moments separately and do not explicitly model complex moment-wise temporal relations. In this paper, we present Moment Alignment Network (MAN), a novel framework that unifies the candidate moment encoding and temporal structural reasoning in a single-shot feed-forward network. MAN naturally assigns candidate moment representations aligned with language semantics over different temporal locations and scales. Most importantly, we propose to explicitly model moment-wise temporal relations as a structured graph and devise an iterative graph adjustment network to jointly learn the best structure in an end-to-end manner. We evaluate the proposed approach on two challenging public benchmarks DiDeMo and Charades-STA, where our MAN significantly outperforms the state-of-the-art by a large margin.

A method called, sigma-consensus, is proposed to eliminate the need for a user-defined inlier-outlier threshold in RANSAC. Instead of estimating the noise sigma, it is marginalized over a range of noise scales. The optimized model is obtained by weighted least-squares fitting where the weights come from the marginalization over sigma of the point likelihoods of being inliers. A new quality function is proposed not requiring sigma and, thus, a set of inliers to determine the model quality. Also, a new termination criterion for RANSAC is built on the proposed marginalization approach. Applying sigma-consensus, MAGSAC is proposed with no need for a user-defined sigma and improving the accuracy of robust estimation significantly. It is superior to the state-of-the-art in terms of geometric accuracy on publicly available real-world datasets for epipolar geometry (F and E) and homography estimation. In addition, applying sigma-consensus only once as a post-processing step to the RANSAC output always improved the model quality on a wide range of vision problems without noticeable deterioration in processing time, adding a few milliseconds.

This work studies the low-rank tensor completion problem, which aims to exactly recover a low-rank tensor from partially observed entries. Our model is inspired by the recently proposed tensor-tensor product (t-product) based on any invertible linear transforms. When the linear transforms satisfy certain conditions, we deduce the new tensor tubal rank, tensor spectral norm, and tensor nuclear norm. Equipped with the tensor nuclear norm, we then solve the tensor completion problem by solving a convex program and provide the theoretical bound for the exact recovery under certain tensor incoherence conditions. The achieved sampling complexity is order-wise optimal. Our model and result greatly extend existing results in the low-rank matrix and tensor completion. Numerical experiments verify our results and the application on image recovery demonstrates the superiority of our method.

Sparse representation based methods have successfully put forward a general framework for robust face recognition through linear reconstruction and sparsity constraints. However, residual modeling in existing works is not yet robust enough when dealing with dense noise. In this paper, we aim at recognizing identities from faces with varying levels of noises of various forms such as occlusion, pixel corruption, or disguise, and take improving the fitting ability of the error model as the key to addressing this problem. To fully capture the characteristics of different noises, we propose a mixed model combining robust sparsity constraint and low-rank constraint, which can deal with random errors and structured errors simultaneously. For random noises such as pixel corruption, we adopt a Laplacian-uniform mixed function for fitting the error distribution. For structured errors like continuous occlusion or disguise, we utilize robust nuclear norm to constrain the rank of the error matrix. An effective iterative reweighted algorithm is then developed to solve the proposed model. Comprehensive experiments were conducted on several benchmark databases for robust face recognition, and the overall results demonstrate that our model is most robust against various kinds of noises, when compared with state-of-the-art methods.

Learning subtle yet discriminative features (e.g., beak and eyes for a bird) plays a significant role in fine-grained image recognition. Existing attention-based approaches localize and amplify significant parts to learn fine-grained details, which often suffer from a limited number of parts and heavy computational cost. In this paper, we propose to learn such fine-grained features from hundreds of part proposals by Trilinear Attention Sampling Network (TASN) in an efficient teacher-student manner. Specifically, TASN consists of 1) a trilinear attention module, which generates attention maps by modeling the inter-channel relationships, 2) an attention-based sampler which highlights attended parts with high resolution, and 3) a feature distiller, which distills part features into an object-level feature by weight sharing and feature preserving strategies. Extensive experiments verify that TASN yields the best performance under the same settings with the most competitive approaches, in iNaturalist-2017, CUB-Bird, and Stanford-Cars datasets.

Previous scene text detection methods have progressed substantially over the past years. However, limited by the receptive field of CNNs and the simple representations like rectangle bounding box or quadrangle adopted to describe text, previous methods may fall short when dealing with more challenging text instances, such as extremely long text and arbitrarily shaped text. To address these two problems, we present a novel text detector namely LOMO, which localizes the text progressively for multiple times (or in other word, LOok More than Once). LOMO consists of a direct regressor (DR), an iterative refinement module (IRM) and a shape expression module (SEM). At first, text proposals in the form of quadrangle are generated by DR branch. Next, IRM progressively perceives the entire long text by iterative refinement based on the extracted feature blocks of preliminary proposals. Finally, a SEM is introduced to reconstruct more precise representation of irregular text by considering the geometry properties of text instance, including text region, text center line and border offsets. The state-of-the-art results on several public benchmarks including ICDAR2017-RCTW, SCUT-CTW1500, Total-Text, ICDAR2015 and ICDAR17-MLT confirm the striking robustness and effectiveness of LOMO.

Most existing attention-based methods on image captioning focus on the current word and visual information in one time step and generate the next word, without considering the visual and linguistic coherence. We propose Look Back (LB) method to embed visual information from the past and Predict Forward (PF) approach to look into future. LB method introduces attention value from the previous time step into the current attention generation to suit visual coherence of human. PF model predicts the next two words in one time step and jointly employs their probabilities for inference. Then the two approaches are combined together as LBPF to further integrate visual information from the past and linguistic information in the future to improve image captioning performance. All the three methods are applied on a classic base decoder, and show remarkable improvements on MSCOCO dataset with small increments on parameter counts. Our LBPF model achieves BLEU-4 / CIDEr / SPICE scores of 37.4 / 116.4 / 21.2 with cross-entropy loss and 38.3 / 127.6 / 22.0 with CIDEr optimization. Our three proposed methods can be easily applied on most attention-based encoder-decoder models for image captioning.

We propose a new learning paradigm, Local to Global Learning (LGL), for Deep Neural Networks (DNNs) to improve the performance of classification problems. The core of LGL is to learn a DNN model from fewer categories (local) to more categories (global) gradually within the entire training set. LGL is most related to the Self-Paced Learning (SPL) algorithm but its formulation is different from SPL. SPL trains its data from simple to complex, while LGL from local to global. In this paper, we incorporate the idea of LGL into the learning objective of DNNs and explain why LGL works better from an information-theoretic perspective. Experiments on the toy data, CIFAR-10, CIFAR-100, and ImageNet dataset show that LGL outperforms the baseline and SPL-based algorithms.

Fine-grained temporal action parsing is important in many applications, such as daily activity understanding, human motion analysis, surgical robotics and others requiring subtle and precise operations over a long-term period. In this paper we propose a novel bilinear pooling operation, which is used in intermediate layers of a temporal convolutional encoder-decoder net. In contrast to previous work, our proposed bilinear pooling is learnable and hence can capture more complex local statistics than the conventional counterpart. In addition, we introduce exact lower-dimension representations of our bilinear forms, so that the dimensionality is reduced without suffering from information loss nor requiring extra computation. We perform extensive experiments to quantitatively analyze our model and show the superior performances to other state-of-the-art pooling work on various datasets.

Encoding individual facial expressions via action units (AUs) coded by the Facial Action Coding System (FACS) has been found to be an effective approach in resolving the ambiguity issue among different expressions. While a number of methods have been proposed for AU detection, robust AU detection in the wild remains a challenging problem because of the diverse baseline AU intensities across individual subjects, and the weakness of appearance signal of AUs. To resolve these issues, in this work, we propose a novel AU detection method by utilizing local information and the relationship of individual local face regions. Through such a local relationship learning, we expect to utilize rich local information to improve the AU detection robustness against the potential perceptual inconsistency of individual local regions. In addition, considering the diversity in the baseline AU intensities of individual subjects, we further regularize local relationship learning via person-specific face shape information, i.e., reducing the influence of person-specific shape information, and obtaining more AU discriminative features. The proposed approach outperforms the state-of-the-art methods on two widely used AU detection datasets in the public domain (BP4D and DISFA).

We propose a novel method of deep spatial matching (DSM) for image retrieval. Initial ranking is based on image descriptors extracted from convolutional neural network activations by global pooling, as in recent state-of-the-art work. However, the same sparse 3D activation tensor is also approximated by a collection of local features. These local features are then robustly matched to approximate the optimal alignment of the tensors. This happens without any network modification, additional layers or training. No local feature detection happens on the original image. No local feature descriptors and no visual vocabulary are needed throughout the whole process. We experimentally show that the proposed method achieves the state-of-the-art performance on standard benchmarks across different network architectures and different global pooling methods. The highest gain in performance is achieved when diffusion on the nearest-neighbor graph of global descriptors is initiated from spatially verified images.

Stereo occlusion boundaries are one-dimensional structures in the visual field that separate foreground regions of a scene that are visible to both eyes (binocular regions) from background regions of a scene that are visible to only one eye (monocular regions). Stereo occlusion boundaries often coincide with object boundaries, and localizing them is useful for tasks like grasping, manipulation, and navigation. This paper describes the local signatures for stereo occlusion boundaries that exist in a stereo cost volume, and it introduces a local detector for them based on a simple feedforward network with relatively small receptive fields. The local detector produces better boundaries than many other stereo methods, even without incorporating explicit stereo matching, top-down contextual cues, or single-image boundary cues based on texture and intensity.

LiveSketch is a novel algorithm for searching large image collections using hand-sketched queries. LiveSketch tackles the inherent ambiguity of sketch search by creating visual suggestions that augment the query as it is drawn, making query specification an iterative rather than one-shot process that helps disambiguate users' search intent. Our technical contributions are: a triplet convnet architecture that incorporates an RNN based variational autoencoder to search for images using vector (stroke-based) queries; real-time clustering to identify likely search intents (and so, targets within the search embedding); and the use of backpropagation from those targets to perturb the input stroke sequence, so suggesting alterations to the query in order to guide the search. We show improvements in accuracy and time-to-task over contemporary baselines using a 67M image corpus.

Visual-to-auditory sensory substitution devices can assist the blind in sensing the visual environment by translating the visual information into a sound pattern. To improve the translation quality, the task performances of the blind are usually employed to evaluate different encoding schemes. In contrast to the toilsome human-based assessment, we argue that machine model can be also developed for evaluation, and more efficient. To this end, we firstly propose two distinct cross-modal perception model w.r.t. the late-blind and congenitally-blind cases, which aim to generate concrete visual contents based on the translated sound. To validate the functionality of proposed models, two novel optimization strategies w.r.t. the primary encoding scheme are presented. Further, we conduct sets of human-based experiments to evaluate and compare them with the conducted machine-based assessments in the cross-modal generation task. Their highly consistent results w.r.t. different encoding schemes indicate that using machine model to accelerate optimization evaluation and reduce experimental cost is feasible to some extent, which could dramatically promote the upgrading of encoding scheme then help the blind to improve their visual perception ability.

In this paper, we present an accurate and scalable approach to the face clustering task. We aim at grouping a set of faces by their potential identities. We formulate this task as a link prediction problem: a link exists between two faces if they are of the same identity. The key idea is that we find the local context in the feature space around an instance (face) contains rich information about the linkage relationship between this instance and its neighbors. By constructing sub-graphs around each instance as input data, which depict the local context, we utilize the graph convolution network (GCN) to perform reasoning and infer the likelihood of linkage between pairs in the sub-graphs. Experiments show that our method is more robust to the complex distribution of faces than conventional methods, yielding favorably comparable results to state-of-the-art methods on standard face clustering benchmarks, and is scalable to large datasets. Furthermore, we show that the proposed method does not need the number of clusters as prior, is aware of noises and outliers, and can be extended to a multi-view version for more accurate clustering accuracy.

We develop Light Field Messaging (LFM), a process of embedding, transmitting, and receiving hidden information in video that is displayed on a screen and captured by a handheld camera. The goal of the system is to minimize perceived visual artifacts of the message embedding, while simultaneously maximizing the accuracy of message recovery on the camera side. LFM requires photographic steganography for embedding messages that can be displayed and camera-captured. Unlike digital steganography, the embedding requirements are significantly more challenging due to the combined effect of the screen's radiometric emittance function, the camera's sensitivity function, and the camera-display relative geometry. We devise and train a network to jointly learn a deep embedding and recovery algorithm that requires no multi-frame synchronization. A key novel component is the camera display transfer function (CDTF) to model the camera-display pipeline. To learn this CDTF we introduce a dataset (Camera-Display 1M) of 1,000,000 camera-captured images collected from 25 camera-display pairs. The result of this work is a high-performance real-time LFM system using consumer-grade displays and smartphone cameras.

Compared with model architectures, the training process, which is also crucial to the success of detectors, has received relatively less attention in object detection. In this work, we carefully revisit the standard training practice of detectors, and find that the detection performance is often limited by the imbalance during the training process, which generally consists in three levels - sample level, feature level, and objective level. To mitigate the adverse effects caused thereby, we propose Libra R-CNN, a simple but effective framework towards balanced learning for object detection. It integrates three novel components: IoU-balanced sampling, balanced feature pyramid, and balanced L1 loss, respectively for reducing the imbalance at sample, feature, and objective level. Benefitted from the overall balanced design, Libra R-CNN significantly improves the detection performance. Without bells and whistles, it achieves 2.5 points and 2.0 points higher Average Precision (AP) than FPN Faster R-CNN and RetinaNet respectively on MSCOCO.

Feature detectors and descriptors are key low-level vision tools that many higher-level tasks build on. Unfortunately these fail in the presence of challenging light transport effects including partial occlusion, low contrast, and reflective or refractive surfaces. Building on spatio-angular imaging modalities offered by emerging light field cameras, we introduce a new and computationally efficient 4D light field feature detector and descriptor: LiFF. LiFF is scale invariant and utilizes the full 4D light field to detect features that are robust to changes in perspective. This is particularly useful for structure from motion (SfM) and other tasks that match features across viewpoints of a scene. We demonstrate significantly improved 3D reconstructions via SfM when using LiFF instead of the leading 2D or 4D features, and show that LiFF runs an order of magnitude faster than the leading 4D approach. Finally, LiFF inherently estimates depth for each feature, opening a path for future research in light field-based SfM.