CVPR 2021 论文列表

The inductive bias of a neural network is largely determined by the architecture and the training algorithm. To achieve good generalization, how to effectively train a neural network is of great importance. We propose a novel orthogonal over-parameterized training (OPT) framework that can provably minimize the hyperspherical energy which characterizes the diversity of neurons on a hypersphere. By maintaining the minimum hyperspherical energy during training, OPT can greatly improve the empirical generalization. Specifically, OPT fixes the randomly initialized weights of the neurons and learns an orthogonal transformation that applies to these neurons. We consider multiple ways to learn such an orthogonal transformation, including unrolling orthogonalization algorithms, applying orthogonal parameterization, and designing orthogonality-preserving gradient descent. For better scalability, we propose the stochastic OPT which performs orthogonal transformation stochastically for partial dimensions of neurons. Interestingly, OPT reveals that learning a proper coordinate system for neurons is crucial to generalization. We provide some insights on why OPT yields better generalization. Extensive experiments validate the superiority of OPT over the standard training.

We study the problem of quantizing N sorted, scalar datapoints with a fixed codebook containing K entries that are allowed to be rescaled. The problem is defined as finding the optimal scaling factor \alpha and the datapoint assignments into the \alpha-scaled codebook to minimize the squared error between original and quantized points. Previously, the globally optimal algorithms for this problem were derived only for certain codebooks (binary and ternary) or under the assumption of certain distributions (Gaussian, Laplacian). By studying the properties of the optimal quantizer, we derive an \calO(NK \log K) algorithm that is guaranteed to find the optimal quantization parameters for any fixed codebook regardless of data distribution. We apply our algorithm to synthetic and real-world neural network quantization problems and demonstrate the effectiveness of our approach.

Deep Neural Networks(DNNs) require huge GPU memory when training on modern image/video databases. Unfortunately, the GPU memory is physically finite, which limits the image resolutions and batch sizes that could be used in training for better DNN performance. Unlike solutions that require physically upgrade GPUs, the Gradient CheckPointing(GCP) training trades computation for more memory beyond existing GPU hardware. GCP only stores a subset of intermediate tensors, called Gradient Checkpoints (GCs), during forward. Then during backward, extra local forwards are conducted to compute the missing tensors. The total training memory cost becomes the sum of (1) the memory cost of the gradient checkpoints and (2) the maximum memory cost of local forwards. To achieve maximal memory cut-offs, one needs optimal algorithms to select GCs. Existing GCP approaches rely on either manual input of GCs or heuristics-based GC search on Linear Computation Graphs (LCGs), and cannot apply to Arbitrary Computation Graphs(ACGs). In this paper, we present theories and optimal algorithms on GC selection that, for the first time, are applicable to ACGs and achieve the maximal memory cut-offs. Extensive experiments show that our approach not only outperforms existing approaches (only applicable on LCGs), and is applicable to a vast family of LCG and ACG networks, such as Alexnet, VGG, ResNet, Densenet, Inception Net and highly complicated DNNs by Network Architecture Search. Our work enables GCP training on ACGs, and cuts off up-to 80% of training memory with a moderate time overhead ( 30%-50%). Codes are available

We propose a novel framework for creating large-scale photorealistic datasets of indoor scenes, with ground truth geometry, material, lighting and semantics. Our goal is to make the dataset creation process widely accessible, allowing researchers to transform scans into datasets with highquality ground truth. We demonstrate our framework by creating a photorealistic synthetic version of the publicly available ScanNet dataset with consistent layout, semantic labels, high quality spatially-varying BRDF and complex lighting. We render photorealistic images, as well as complex spatially-varying lighting, including direct, indirect and visibility components. Such a dataset enables important applications in inverse rendering, scene understanding and robotics. We show that deep networks trained on the proposed dataset achieve competitive performance for shape, material and lighting estimation on real images, enabling photorealistic augmented reality applications, such as object insertion and material editing. We also show our semantic labels may be used for segmentation and multitask learning. Finally, we demonstrate that our framework may also be integrated with physics engines, to create virtual robotics environments with unique ground truth such as friction coefficients and correspondence to real scenes. The dataset and all the tools to create such datasets will be publicly released, enabling others in the community to easily build large-scale datasets of their own.

In this paper, we tackle the problem of discovering new classes in unlabeled visual data given labeled data from disjoint classes. Existing methods typically first pre-train a model with labeled data, and then identify new classes in unlabeled data via unsupervised clustering. However, the labeled data that provide essential knowledge are often underexplored in the second step. The challenge is that the labeled and unlabeled examples are from non-overlapping classes, which makes it difficult to build a learning relationship between them. In this work, we introduce OpenMix to mix the unlabeled examples from an open set and the labeled examples from known classes, where their non-overlapping labels and pseudo-labels are simultaneously mixed into a joint label distribution. OpenMix dynamically compounds examples in two ways. First, we produce mixed training images by incorporating labeled examples with unlabeled examples. With the benefit of unique prior knowledge in novel class discovery, the generated pseudo-labels will be more credible than the original unlabeled predictions. As a result, OpenMix helps preventing the model from overfitting on unlabeled samples that may be assigned with wrong pseudo-labels. Second, the first way encourages the unlabeled examples with high class-probabilities to have considerable accuracy. We introduce these examples as reliable anchors and further integrate them with unlabeled samples. This enables us to generate more combinations in unlabeled examples and exploit finer object relations among the new classes. Experiments on three classification datasets demonstrate the effectiveness of the proposed OpenMix, which is superior to state-of-the-art methods in novel class discovery.

Despite the remarkable accuracy of deep neural networks in object detection, they are costly to train and scale due to supervision requirements. Particularly, learning more object categories typically requires proportionally more bounding box annotations. Weakly supervised and zero-shot learning techniques have been explored to scale object detectors to more categories with less supervision, but they have not been as successful and widely adopted as supervised models. In this paper, we put forth a novel formulation of the object detection problem, namely open-vocabulary object detection, which is more general, more practical, and more effective than weakly supervised and zero-shot approaches. We propose a new method to train object detectors using bounding box annotations for a limited set of object categories, as well as image-caption pairs that cover a larger variety of objects at a significantly lower cost. We show that the proposed method can detect and localize objects for which no bounding box annotation is provided during training, at a significantly higher accuracy than zero-shot approaches. Meanwhile, objects with bounding box annotation can be detected almost as accurately as supervised methods, which is significantly better than weakly supervised baselines. Accordingly, we establish a new state of the art for scalable object detection.

In this paper, we convert traditional video captioning task into a new paradigm, i.e., Open-book Video Captioning, which generates natural language under the prompts of video-content-relevant sentences, not limited to the video itself. To address the open-book video captioning problem, we propose a novel Retrieve-Copy-Generate network, where a pluggable video-to-text retriever is leveraged to effectively retrieve sentences as hints from the training corpus, and a copy-mechanism generator is introduced to dynamically extract expressions from multi-retrievals. The two modules can be trained end-to-end or separately which is flexible and extensible. Our framework coordinates the conventional retrieval based methods with orthodox encoder-decoder methods, which can not only draw on the diverse expressions in the retrieved sentences but also generate natural and accurate content of the video. Extensive experiments on several benchmark datasets show that our proposed approach performs better than state-of-the-art approaches, indicating the effectiveness and promising of the proposed paradigm in the task of video captioning.

Compositional Zero-Shot learning (CZSL) requires to recognize state-object compositions unseen during training. In this work, instead of assuming prior knowledge about the unseen compositions, we operate in the open world setting, where the search space includes a large number of unseen compositions some of which might be unfeasible. In this setting, we start from the cosine similarity between visual features and compositional embeddings. After estimating the feasibility score of each composition, we use these scores to either directly mask the output space or as a margin for the cosine similarity between visual features and compositional embeddings during training. Our experiments on two standard CZSL benchmarks show that all the methods suffer severe performance degradation when applied in the open world setting. While our simple CZSL model achieves state-of-the-art performances in the closed world scenario, our feasibility scores boost the performance of our approach in the open world setting, clearly outperforming the previous state of the art. Code is available at: https://github.com/ExplainableML/czsl.

Leveraging datasets available to learn a model with high generalization ability to unseen domains is important for computer vision, especially when the unseen domain's annotated data are unavailable. We study the problem of learning from different source domains to achieve high performance on an unknown target domain, where the distributions and label sets of each individual source domain and the target domain are different. The problem can be generally applied to diverse source domains and widely applicable to real-world applications. We propose a Domain-Augmented Meta-Learning framework to learn open-domain generalizable representations. We augment domains on both feature-level by a new Dirichlet mixup and label-level by distilled soft-labeling, which complements each domain with missing classes and other domain knowledge. We conduct meta-learning over domains by designing new meta-learning tasks and losses to preserve domain unique knowledge and generalize knowledge across domains simultaneously. Experiment results on various multi-domain datasets demonstrate that the proposed Domain-Augmented Meta-Learning outperforms previous methods for unseen target classification.

Modern online multiple object tracking (MOT) methods usually focus on two directions to improve tracking performance. One is to predict new positions in an incoming frame based on tracking information from previous frames, and the other is to enhance data association by generating more discriminative identity embeddings. Some works combined both directions within one framework but handled them as two individual tasks, thus gaining little mutual benefits. In this paper, we propose a novel unified model with synergy between position prediction and embedding association. The two tasks are linked by temporal-aware target attention and distractor attention, as well as identity-aware memory aggregation model. Specifically, the attention modules can make the prediction focus more on targets and less on distractors, therefore more reliable embeddings can be extracted accordingly for association. On the other hand, such reliable embeddings can boost identity-awareness through memory aggregation, hence strengthen attention modules and suppress drifts. In this way, the synergy between position prediction and embedding association is achieved, which leads to strong robustness to occlusions. Extensive experiments demonstrate the superiority of our proposed model over a wide range of existing methods on MOTChallenge benchmarks. Our code and models are publicly available at https://github.com/songguocode/TADAM

Constructing and maintaining a consistent scene model on-the-fly is the core task for online spatial perception, interpretation, and action. In this paper, we represent the scene with a Bayesian nonparametric mixture model, seamlessly describing per-point occupancy status with a continuous probability density function. Instead of following the conventional data fusion paradigm, we address the problem of online learning the process how sequential point cloud data are generated from the scene geometry. An incremental and parallel inference is performed to update the parameter space in real-time. We experimentally show that the proposed representation achieves state-of-the-art accuracy with promising efficiency. The consistent probabilistic formulation assures a generative model that is adaptive to different sensor characteristics, and the model complexity can be dynamically adjusted on-the-fly according to different data scales.

Despite remarkable progress achieved, most neural architecture search (NAS) methods focus on searching for one single accurate and robust architecture. To further build models with better generalization capability and performance, model ensemble is usually adopted and performs better than stand-alone models. Inspired by the merits of model ensemble, we propose to search for multiple diverse models simultaneously as an alternative way to find powerful models. Searching for ensembles is non-trivial and has two key challenges: enlarged search space and potentially more complexity for the searched model. In this paper, we propose a one-shot neural ensemble architecture search (NEAS) solution that addresses the two challenges. For the first challenge, we introduce a novel diversity-based metric to guide search space shrinking, considering both the potentiality and diversity of candidate operators. For the second challenge, we enable a new search dimension to learn layer sharing among different models for efficiency purposes. The experiments on ImageNet clearly demonstrate that our solution can improve the supernet's capacity of ranking ensemble architectures, and further lead to better search results. The discovered architectures achieve superior performance compared with state-of-the-arts such as MobileNetV3 and EfficientNet families under aligned settings. Moreover, we evaluate the generalization ability and robustness of our searched architecture on the COCO detection benchmark and achieve a 3.1% improvement on AP compared with MobileNetV3. Codes and models are available here.

We propose a neural talking-head video synthesis model and demonstrate its application to video conferencing. Our model learns to synthesize a talking-head video using a source image containing the target person's appearance and a driving video that dictates the motion in the output. Our motion is encoded based on a novel keypoint representation, where the identity-specific and motion-related information is decomposed unsupervisedly. Extensive experimental validation shows that our model outperforms competing methods on benchmark datasets. Moreover, our compact keypoint representation enables a video conferencing system that achieves the same visual quality as the commercial H.264 standard while only using one-tenth of the bandwidth. Besides, we show our keypoint representation allows the user to rotate the head during synthesis, which is useful for simulating face-to-face video conferencing experiences.

Point cloud semantic segmentation often requires largescale annotated training data, but clearly, point-wise labels are too tedious to prepare. While some recent methods propose to train a 3D network with small percentages of point labels, we take the approach to an extreme and propose "One Thing One Click," meaning that the annotator only needs to label one point per object. To leverage these extremely sparse labels in network training, we design a novel self-training approach, in which we iteratively conduct the training and label propagation, facilitated by a graph propagation module. Also, we adopt a relation network to generate per-category prototype and explicitly model the similarity among graph nodes to generate pseudo labels to guide the iterative training. Experimental results on both ScanNet-v2 and S3DIS show that our self-training approach, with extremely-sparse annotations, outperforms all existing weakly supervised methods for 3D semantic segmentation by a large margin, and our results are also comparable to those of the fully supervised counterparts.

Face swapping has both positive applications such as entertainment, human-computer interaction, etc., and negative applications such as DeepFake threats to politics, economics, etc. Nevertheless, it is necessary to understand the scheme of advanced methods for high-quality face swapping and generate enough and representative face swapping images to train DeepFake detection algorithms. This paper proposes the first Megapixel level method for one shot Face Swapping (or MegaFS for short). Firstly, MegaFS organizes face representation hierarchically by the proposed Hierarchical Representation Face Encoder (HieRFE) in an extended latent space to maintain more facial details, rather than compressed representation in previous face swapping methods. Secondly, a carefully designed Face Transfer Module (FTM) is proposed to transfer the identity from a source image to the target by a non-linear trajectory without explicit feature disentanglement. Finally, the swapped faces can be synthesized by StyleGAN2 with the benefits of its training stability and powerful generative capability. Each part of MegaFS can be trained separately so the requirement of our model for GPU memory can be satisfied for megapixel face swapping. In summary, complete face representation, stable training, and limited memory usage are the three novel contributions to the success of our method. Extensive experiments demonstrate the superiority of MegaFS and the first megapixel level face swapping database is released for research on DeepFake detection and face image editing in the public domain.

Recent studies propose membership inference (MI) attacks on deep models, where the goal is to infer if a sample has been used in the training process. Despite their apparent success, these studies only report accuracy, precision, and recall of the positive class (member class). Hence, the performance of these attacks have not been clearly reported on negative class (non-member class). In this paper, we show that the way the MI attack performance has been reported is often misleading because they suffer from high false positive rate or false alarm rate (FAR) that has not been reported. FAR shows how often the attack model mislabel non-training samples (non-member) as training (member) ones. The high FAR makes MI attacks fundamentally impractical, which is particularly more significant for tasks such as membership inference where the majority of samples in reality belong to the negative (non-training) class. Moreover, we show that the current MI attack models can only identify the membership of misclassified samples with mediocre accuracy at best, which only constitute a very small portion of training samples. We analyze several new features that have not been comprehensively explored for membership inference before, including distance to the decision boundary and gradient norms, and conclude that deep models' responses are mostly similar among train and non-train samples. We conduct several experiments on image classification tasks, including MNIST, CIFAR-10, CIFAR-100, and ImageNet, using various model architecture, including LeNet, AlexNet, ResNet, etc. We show that the current state-of-the-art MI attacks cannot achieve high accuracy and low FAR at the same time, even when the attacker is given several advantages. The source code is available at https://github.com/shrezaei/MI-Attack.

Current video retrieval efforts all found their evaluation on an instance-based assumption, that only a single caption is relevant to a query video and vice versa. We demonstrate that this assumption results in performance comparisons often not indicative of models' retrieval capabilities. We propose a move to semantic similarity video retrieval, where (i) multiple videos/captions can be deemed equally relevant, and their relative ranking does not affect a method's reported performance and (ii) retrieved videos/captions are ranked by their similarity to a query. We propose several proxies to estimate semantic similarities in large-scale retrieval datasets, without additional annotations. Our analysis is performed on three commonly used video retrieval datasets (MSR-VTT, YouCook2 and EPIC-KITCHENS).

People touch their face 23 times an hour, they cross their arms and legs, put their hands on their hips, etc. While many images of people contain some form of self-contact, current 3D human pose and shape (HPS) regression methods typically fail to estimate this contact. To address this, we develop new datasets and methods that significantly improve human pose estimation with self-contact. First, we create a dataset of 3D Contact Poses (3DCP) containing SMPL-X bodies fit to 3D scans as well as poses from AMASS, which we refine to ensure good contact. Second, we leverage this to create the Mimic-The-Pose (MTP) dataset of images, collected via Amazon Mechanical Turk, containing people mimicking the 3DCP poses with self-contact. Third, we develop a novel HPS optimization method, SMPLify-XMC, that includes contact constraints and uses the known 3DCP body pose during fitting to create near ground-truth poses for MTP images. Fourth, for more image variety, we label a dataset of in-the-wild images with Discrete Self-Contact (DSC) information and use another new optimization method, SMPLify-DC, that exploits discrete contacts during pose optimization. Finally, we use our datasets during SPIN training to learn a new 3D human pose regressor, called TUCH (Towards Understanding Contact in Humans). We show that the new self-contact training data significantly improves 3D human pose estimates on withheld test data and existing datasets like 3DPW. Not only does our method improve results for self-contact poses, but it also improves accuracy for non-contact poses. The code and data are available for research purposes at https://tuch.is.tue.mpg.de.

Modern deep convolutional networks (CNNs) are often criticized for not generalizing under distributional shifts. However, several recent breakthroughs in transfer learning suggest that these networks can cope with severe distribution shifts and successfully adapt to new tasks from a few training examples. In this work we study the interplay between out-of-distribution and transfer performance of modern image classification CNNs for the first time and investigate the impact of the pre-training data size, the model scale, and the data preprocessing pipeline. We find that increasing both the training set and model sizes significantly improve the distributional shift robustness. Furthermore, we show that, perhaps surprisingly, simple changes in the preprocessing such as modifying the image resolution can significantly mitigate robustness issues in some cases. Finally, we outline the shortcomings of existing robustness evaluation datasets and introduce a synthetic dataset SI-Score we use for a systematic analysis across factors of variation common in visual data such as object scale and position.