CVPR 2021 论文列表

Knowledge transfer from large teacher models to smaller student models has recently been studied for metric learning, focusing on fine-grained classification. In this work, focusing on instance-level image retrieval, we study an asymmetric testing task, where the database is represented by the teacher and queries by the student. Inspired by this task, we introduce asymmetric metric learning, a novel paradigm of using asymmetric representations at training. This acts as a simple combination of knowledge transfer with the original metric learning task. We systematically evaluate different teacher and student models, metric learning and knowledge transfer loss functions on the new asymmetric testing as well as the standard symmetric testing task, where database and queries are represented by the same model. We find that plain regression is surprisingly effective compared to more complex knowledge transfer mechanisms, working best in asymmetric testing. Interestingly, our asymmetric metric learning approach works best in symmetric testing, allowing the student to even outperform the teacher. Our implementation is publicly available, including trained student models for all loss functions and all pairs of teacher/student models. This can serve as a benchmark for future research.

With the development of deep learning techniques, the combination of deep learning with image compression has drawn lots of attention. Recently, learned image compression methods had exceeded their classical counterparts in terms of rate-distortion performance. However, continuous rate adaptation remains an open question. Some learned image compression methods use multiple networks for multiple rates, while others use one single model at the expense of computational complexity increase and performance degradation. In this paper, we propose a continuously rate adjustable learned image compression framework, Asymmetric Gained Variational Autoencoder (AG-VAE). AG-VAE utilizes a pair of gain units to achieve discrete rate adaptation in one single model with a negligible additional computation. Then, by using exponential interpolation, continuous rate adaptation is achieved without compromising performance. Besides, we propose the asymmetric Gaussian entropy model for more accurate entropy estimation. Exhaustive experiments show that our method achieves comparable quantitative performance with SOTA learned image compression methods and better qualitative performance than classical image codecs. In the ablation study, we confirm the usefulness and superiority of gain units and the asymmetric Gaussian entropy model.

Universal style transfer retains styles from reference images in content images. While existing methods have achieved state-of-the-art style transfer performance, they are not aware of the content leak phenomenon that the image content may corrupt after several rounds of stylization process. In this paper, we propose ArtFlow to prevent content leak during universal style transfer. ArtFlow consists of reversible neural flows and an unbiased feature transfer module. It supports both forward and backward inferences and operates in a projection-transfer-reversion scheme. The forward inference projects input images into deep features, while the backward inference remaps deep features back to input images in a lossless and unbiased way. Extensive experiments demonstrate that ArtFlow achieves comparable performance to state-of-the-art style transfer methods while avoiding content leak.

We present a novel large-scale dataset and accompanying machine learning models aimed at providing a detailed understanding of the interplay between visual content, its emotional effect, and explanations for the latter in language. In contrast to most existing annotation datasets in computer vision, we focus on the affective experience triggered by visual artworks and ask the annotators to indicate the dominant emotion they feel for a given image and, crucially, to also provide a grounded verbal explanation for their emotion choice. As we demonstrate below, this leads to a rich set of signals for both the objective content and the affective impact of an image, creating associations with abstract concepts (e.g., "freedom" or "love"), or references that go beyond what is directly visible, including visual similes and metaphors, or subjective references to personal experiences. We focus on visual art (e.g., paintings, artistic photographs) as it is a prime example of imagery created to elicit emotional responses from its viewers. Our dataset, termed ArtEmis, contains 455K emotion attributions and explanations from humans, on 80K artworks from WikiArt. Building on this data, we train and demonstrate a series of captioning systems capable of expressing and explaining emotions from visual stimuli. Remarkably, the captions produced by these systems often succeed in reflecting the semantic and abstract content of the image, going well beyond systems trained on existing datasets.

Quick Response (QR) code is one of the most worldwide used two-dimensional codes. Traditional QR codes appear as random collections of black-and-white modules that lack visual semantics and aesthetic elements, which inspires the recent works to beautify the appearances of QR codes. However, these works typically beatify QR codes in a single style due to the fixed generation algorithms, which is improvable in personalization and diversification. In this paper, combining the Neural Style Transfer technique, we propose a novel end-to-end network ACN (ArtCode-Net) to generate the stylized QR codes that are personalized, diverse, attractive, and scanning-robust. To address the challenge that preserving the scanning-robustness after giving such codes style elements, we further propose the Sampling-Simulation layer, the module-based code loss, and a competition mechanism to improve the performances of ACN. The experimental results show that our stylized QR codes have high-quality in both the visual effect and the scanning-robustness, and they are able to support the real-world application.

To calculate the model accuracy on a computer vision task, e.g., object recognition, we usually require a test set composing of test samples and their ground truth labels. Whilst standard usage cases satisfy this requirement, many real-world scenarios involve unlabeled test data, rendering common model evaluation methods infeasible. We investigate this important and under-explored problem, Automatic model Evaluation (AutoEval). Specifically, given a labeled training set and a classifier, we aim to estimate the classification accuracy on unlabeled test datasets. We construct a meta-dataset: a dataset comprised of datasets generated from the original images via various transformations such as rotation, background substitution, foreground scaling, etc. As the classification accuracy of the model on each sample (dataset) is known from the original dataset labels, our task can be solved via regression. Using the feature statistics to represent the distribution of a sample dataset, we can train regression models (e.g., a regression neural network) to predict model performance. Using synthetic meta-dataset and real-world datasets in training and testing, respectively, we report a reasonable and promising prediction of the model accuracy. We also provide insights into the application scope, limitation, and potential future direction of AutoEval.

Despite their unmatched performance, deep neural networks remain susceptible to targeted attacks by nearly imperceptible levels of adversarial noise. While the underlying cause of this sensitivity is not well understood, theoretical analyses can be simplified by reframing each layer of a feed-forward network as an approximate solution to a sparse coding problem. Iterative solutions using basis pursuit are theoretically more stable and have improved adversarial robustness. However, cascading layer-wise pursuit implementations suffer from error accumulation in deeper networks. In contrast, our new method of deep pursuit approximates the activations of all layers as a single global optimization problem, allowing us to consider deeper, real-world architectures with skip connections such as residual networks. Experimentally, our approach demonstrates improved robustness to adversarial noise.

Generative adversarial networks (GANs) have enabled photorealistic image synthesis and editing. However, due to the high computational cost of large-scale generators (e.g., StyleGAN2), it usually takes seconds to see the results of a single edit on edge devices, prohibiting interactive user experience. In this paper, inspired by quick preview features in modern rendering software, we propose Anycost GAN for interactive natural image editing. We train the Anycost GAN to support elastic resolutions and channels for faster image generation at versatile speeds. Running subsets of the full generator produce outputs that are perceptually similar to the full generator, making them a good proxy for a quick preview. By using sampling-based multi-resolution training, adaptive-channel training, and a generator-conditioned discriminator, the anycost generator can be evaluated at various configurations while achieving better image quality compared to separately trained models. Furthermore, we develop new encoder training and latent code optimization techniques to encourage consistency between the different sub-generators during image projection. Anycost GAN can be executed at various cost budgets (up to 10x computation reduction) and adapt to a wide range of hardware and latency requirements. When deployed on desktop CPUs and edge devices, our model can provide perceptually similar previews at 6-12x speedup, enabling interactive image editing. The code and demo are publicly available.

We propose a framework for early action recognition and anticipation by correlating past features with the future using three novel similarity measures called Jaccard vector similarity, Jaccard cross-correlation and Jaccard Frobenius inner product over covariances. Using these combinations of novel losses and using our framework, we obtain state-of-the-art results for early action recognition in UCF101 and JHMDB datasets by obtaining 91.7 % and 83.5 % accuracy respectively for an observation percentage of 20. Similarly, we obtain state-of-the-art results for Epic-Kitchen55 and Breakfast datasets for action anticipation by obtaining 20.35 and 41.8 top-1 accuracy respectively.

Encouraging progress in few-shot semantic segmentation has been made by leveraging features learned upon base classes with sufficient training data to represent novel classes with few-shot examples. However, this feature sharing mechanism inevitably causes semantic aliasing between novel classes when they have similar compositions of semantic concepts. In this paper, we reformulate few-shot segmentation as a semantic reconstruction problem, and convert base class features into a series of basis vectors which span a class-level semantic space for novel class reconstruction. By introducing contrastive loss, we maximize the orthogonality of basis vectors while minimizing semantic aliasing between classes. Within the reconstructed representation space, we further suppress interference from other classes by projecting query features to the support vector for precise semantic activation. Our proposed approach, referred to as anti-aliasing semantic reconstruction (ASR), provides a systematic yet interpretable solution for few-shot learning problems. Extensive experiments on PASCAL VOC and MS COCO datasets show that ASR achieves strong results compared with the prior works. Code will be released at github.com/Bibkiller/ASR.

Weakly supervised semantic segmentation produces a pixel-level localization from class labels; but a classifier trained on such labels is likely to restrict its focus to a small discriminative region of the target object. AdvCAM is an attribution map of an image that is manipulated to increase the classification score produced by a classifier. This manipulation is realized in an anti-adversarial manner, which perturbs the original images along pixel gradients in the opposite direction from those used in an adversarial attack. It forces regions initially considered not to be discriminative to become involved in subsequent classifications, and produces attribution maps that successively identify more regions of the target object. In addition, we introduce a new regularization procedure that inhibits both the incorrect attribution of regions unrelated to the target object and excessive concentration of attributions on a small region of that object. Our method is a post-hoc analysis of a trained classifier, which does not need to be altered or retrained. On PASCAL VOC 2012 test images, we achieve mIoUs of 68.0 and 76.9 for weakly and semi-supervised semantic segmentation respectively, which represent a new state-of-the-art.

Anomaly detection in video is a challenging computer vision problem. Due to the lack of anomalous events at training time, anomaly detection requires the design of learning methods without full supervision. In this paper, we approach anomalous event detection in video through self-supervised and multi-task learning at the object level. We first utilize a pre-trained detector to detect objects. Then, we train a 3D convolutional neural network to produce discriminative anomaly-specific information by jointly learning multiple proxy tasks: three self-supervised and one based on knowledge distillation. The self-supervised tasks are: (i) discrimination of forward/backward moving objects (arrow of time), (ii) discrimination of objects in consecutive/intermittent frames (motion irregularity) and (iii) reconstruction of object-specific appearance information. The knowledge distillation task takes into account both classification and detection information, generating large prediction discrepancies between teacher and student models when anomalies occur. To the best of our knowledge, we are the first to approach anomalous event detection in video as a multi-task learning problem, integrating multiple self-supervised and knowledge distillation proxy tasks in a single architecture. Our lightweight architecture outperforms the state-of-the-art methods on three benchmarks: Avenue, ShanghaiTech and UCSD Ped2. Additionally, we perform an ablation study demonstrating the importance of integrating self-supervised learning and normality-specific distillation in a multi-task learning setting.

In this paper, we demonstrate a fully automatic method for converting a still image into a realistic animated looping video. We target scenes with continuous fluid motion, such as flowing water and billowing smoke. Our method relies on the observation that this type of natural motion can be convincingly reproduced from a static Eulerian motion description, i.e. a single, temporally constant flow field that defines the immediate motion of a particle at a given 2D location. We use an image-to-image translation network to encode motion priors of natural scenes collected from online videos, so that for a new photo, we can synthesize a corresponding motion field. The image is then animated using the generated motion through a deep warping technique: pixels are encoded as deep features, those features are warped via Eulerian motion, and the resulting warped feature maps are decoded as images. In order to produce continuous, seamlessly looping video textures, we propose a novel video looping technique that flows features both forward and backward in time and then blends the results. We demonstrate the effectiveness and robustness of our method by applying it to a large collection of examples including beaches, waterfalls, and flowing rivers.

Person search aims to simultaneously localize and identify a query person from realistic, uncropped images, which can be regarded as the unified task of pedestrian detection and person re-identification (re-id). Most existing works employ two-stage detectors like Faster-RCNN, yielding encouraging accuracy but with high computational overhead. In this work, we present the Feature-Aligned Person Search Network (AlignPS), the first anchor-free framework to efficiently tackle this challenging task. AlignPS explicitly addresses the major challenges, which we summarize as the misalignment issues in different levels (i.e., scale, region, and task), when accommodating an anchor-free detector for this task. More specifically, we propose an aligned feature aggregation module to generate more discriminative and robust feature embeddings by following a "re-id first" principle. Such a simple design directly improves the baseline anchor-free model on CUHK-SYSU by more than 20% in mAP. Moreover, AlignPS outperforms state-of-the-art two-stage methods, with a higher speed. The code is available at https://github.com/daodaofr/AlignPS.

This paper is about action segmentation under weak supervision in training, where the ground truth provides only a set of actions present, but neither their temporal ordering nor when they occur in a training video. We use a Hidden Markov Model (HMM) grounded on a multilayer perceptron (MLP) to label video frames, and thus generate a pseudo-ground truth for the subsequent pseudo-supervised training. In testing, a Monte Carlo sampling of action sets seen in training is used to generate candidate temporal sequences of actions, and select the maximum posterior sequence. Our key contribution is a new anchor-constrained Viterbi algorithm (ACV) for generating the pseudo-ground truth, where anchors are salient action parts estimated for each action from a given ground-truth set. Our evaluation on the tasks of action segmentation and alignment on the benchmark Breakfast, MPII Cooking2, Hollywood Extended datasets demonstrates our superior performance relative to that of prior work.

The Huber loss is a robust loss function used for a wide range of regression tasks. To utilize the Huber loss, a parameter that controls the transitions from a quadratic function to an absolute value function needs to be selected. We believe the standard probabilistic interpretation that relates the Huber loss to the Huber density fails to provide adequate intuition for identifying the transition point. As a result, a hyper-parameter search is often necessary to determine an appropriate value. In this work, we propose an alternative probabilistic interpretation of the Huber loss, which relates minimizing the loss to minimizing an upper-bound on the Kullback-Leibler divergence between Laplace distributions, where one distribution represents the noise in the ground-truth and the other represents the noise in the prediction. In addition, we show that the parameters of the Laplace distributions are directly related to the transition point of the Huber loss. We demonstrate, through a toy problem, that the optimal transition point of the Huber loss is closely related to the distribution of the noise in the ground-truth data. As a result, our interpretation provides an intuitive way to identify well-suited hyper-parameters by approximating the amount of noise in the data, which we demonstrate through a case study and experimentation on the Faster R-CNN and RetinaNet object detectors.

In this paper, we study a novel knowledge transfer task in the domain of graph neural networks (GNNs). We strive to train a multi-talented student GNN, without accessing human annotations, that "amalgamates" knowledge from a couple of teacher GNNs with heterogeneous architectures and handling distinct tasks. The student derived in this way is expected to integrate the expertise from both teachers while maintaining a compact architecture. To this end, we propose an innovative approach to train a slimmable GNN that enables learning from teachers with varying feature dimensions. Meanwhile, to explicitly align topological semantics between the student and teachers, we introduce a topological attribution map (TAM) to highlight the structural saliency in a graph, based on which the student imitates the teachers' ways of aggregating information from neighbors. Experiments on seven datasets across various tasks, including multi-label classification and joint segmentation-classification, demonstrate that the learned student, with a lightweight architecture, achieves gratifying results on par with and sometimes even superior to those of the teachers in their specializations. Our code is publicly available at https://github.com/ycjing/AmalgamateGNN.PyTorch.

Semi-supervised learning (SSL) is a key approach toward more data-efficient machine learning by jointly leverage both labeled and unlabeled data. We propose AlphaMatch, an efficient SSL method that leverages data augmentations, by efficiently enforcing the label consistency between the data points and the augmented data derived from them. Our key technical contribution lies on: 1) using alpha-divergence to prioritize the regularization on data with high Semi-supervised learning (SSL) is a key approach toward more data-efficient machine learning by jointly leverage both labeled and unlabeled data. We propose AlphaMatch, an efficient SSL method that leverages data augmentations, by efficiently enforcing the label consistency between the data points and the augmented data derived from them. Our key technical contribution lies on: 1) using alpha-divergence to prioritize the regularization on data with high confidence, achieving similar effect as FixMatch but in a more flexible fashion, and 2) proposing an optimization-based, EM-like algorithm to enforce the consistency, which enjoys better convergence than iterative regularization procedures used in recent SSL methods such as FixMatch, UDA, and MixMatch. AlphaMatch is simple and easy to implement, and consistently outperforms prior arts on standard benchmarks, e.g. CIFAR-10, SVHN, CIFAR-100, STL-10. Specifically, we achieve 91.3% test accuracy on CIFAR-10 with just 4 labelled data per class, substantially improving over the previously best 88.7% accuracy achieved by FixMatch.

Visual object tracking aims to precisely estimate the bounding box for the given target, which is a challenging problem due to factors such as deformation and occlusion. Many recent trackers adopt the multiple-stage tracking strategy to improve the quality of bounding box estimation. These methods first coarsely locate the target and then refine the initial prediction in the following stages. However, existing approaches still suffer from limited precision, and the coupling of different stages severely restricts the method's transferability. This work proposes a novel, flexible, and accurate refinement module called Alpha-Refine (AR), which can significantly improve the base trackers' box estimation quality. By exploring a series of design options, we conclude that the key to successful refinement is extracting and maintaining detailed spatial information as much as possible. Following this principle, Alpha-Refine adopts a pixel-wise correlation, a corner prediction head, and an auxiliary mask head as the core components. Comprehensive experiments on TrackingNet, LaSOT, GOT-10K, and VOT2020 benchmarks with multiple base trackers show that our approach significantly improves the base trackers' performance with little extra latency. The proposed Alpha-Refine method leads to a series of strengthened trackers, among which the ARSiamRPN (AR strengthened SiamRPNpp) and the ARDiMP50 (ARstrengthened DiMP50) achieve good efficiency-precision trade-off, while the ARDiMPsuper (AR strengthened DiMP-super) achieves very competitive performance at a real-time speed. Code and pretrained models are available at https://github.com/MasterBin-IIAU/AlphaRefine.