CVPR 2022 Papers

Annotating images with pixel-wise labels is a time-consuming and costly process. Recently, DatasetGAN showcased a promising alternative - to synthesize a large labeled dataset via a generative adversarial network (GAN) by exploiting a small set of manually labeled, GAN-generated images. Here, we scale DatasetGAN to ImageNet scale of class diversity. We take image samples from the class-conditional generative model BigGAN trained on ImageNet, and manually annotate only 5 images per class, for all 1k classes. By training an effective feature segmentation architecture on top of BigGAN, we turn BigGAN into a labeled dataset generator. We further show that VQGAN can similarly serve as a dataset generator, leveraging the already annotated data. We create a new ImageNet benchmark by labeling an additional set of real images and evaluate segmentation performance in a variety of settings. Through an extensive ablation study we show big gains in leveraging a large generated dataset to train different supervised and self-supervised backbone models on pixel-wise tasks. Furthermore, we demonstrate that using our synthesized datasets for pre-training leads to improvements over standard ImageNet pre-training on several downstream datasets, such as PASCAL-VOC, MS-COCO, Cityscapes and chest X-ray, as well as tasks (detection, segmentation). Our benchmark will be made public and maintain a leaderboard for this challenging task.

Energy-based latent variable models (EBLVMs) are more expressive than conventional energy-based models. However, its potential on visual tasks are limited by its training process based on maximum likelihood estimate that requires sampling from two intractable distributions. In this paper, we propose Bi-level doubly variational learning (BiDVL), which is based on a new bi-level optimization framework and two tractable variational distributions to facilitate learning EBLVMs. Particularly, we lead a decoupled EBLVM consisting of a marginal energy-based distribution and a structural posterior to handle the difficulties when learning deep EBLVMs on images. By choosing a symmetric KL divergence in the lower level of our framework, a compact BiDVL for visual tasks can be obtained. Our model achieves impressive image generation performance over related works. It also demonstrates the significant capacity of testing image reconstruction and out-of-distribution detection.

Recently, crowd density estimation has received increasing attention. The main challenge for this task is to achieve high-quality manual annotations on a large amount of training data. To avoid reliance on such annotations, previous works apply unsupervised domain adaptation (UDA) techniques by transferring knowledge learned from easily accessible synthetic data to real-world datasets. However, current state-of-the-art methods either rely on external data for training an auxiliary task or apply an expensive coarse-to-fine estimation. In this work, we aim to develop a new adversarial learning based method, which is simple and efficient to apply. To reduce the domain gap between the synthetic and real data, we design a bi-level alignment framework (BLA) consisting of (1) task-driven data alignment and (2) fine-grained feature alignment. Contrast to previous domain augmentation methods, we introduce AutoML to search for an optimal transform on source, which well serves for the downstream task. On the other hand, we do fine-grained alignment for foreground and background separately to alleviate the alignment difficulty. We evaluate our approach on five real-world crowd counting benchmarks, where we outperform existing approaches by a large margin. Also, our approach is simple, easy to implement and efficient to apply. The code will be made publicly available. https://github.com/Yankeegsj/BLA

The saliency ranking task is recently proposed to study the visual behavior that humans would typically shift their attention over different objects of a scene based on their degrees of saliency. Existing approaches focus on learning either object-object or object-scene relations. Such a strategy follows the idea of object-based attention in Psychology, but it tends to favor those objects with strong semantics (e.g., humans), resulting in unrealistic saliency ranking. We observe that spatial attention works concurrently with object-based attention in the human visual recognition system. During the recognition process, the human spatial attention mechanism would move, engage, and disengage from region to region (i.e., context to context). This inspires us to model the region-level interactions, in addition to the object-level reasoning, for saliency ranking. To this end, we propose a novel bi-directional method to unify spatial attention and object-based attention for saliency ranking. Our model includes two novel modules: (1) a selective object saliency (SOS) module that models object-based attention via inferring the semantic representation of the salient object, and (2) an object-context-object relation (OCOR) module that allocates saliency ranks to objects by jointly modeling the object-context and context-object interactions of the salient objects. Extensive experiments show that our approach outperforms existing state-of-the-art methods. Code and pretrained model are available at https://github.com/GrassBro/OCOR.

Significant progress has been made on visual captioning, largely relying on pre-trained features and later fixed object detectors that serve as rich inputs to auto-regressive models. A key limitation of such methods, however, is that the output of the model is conditioned only on the object detector's outputs. The assumption that such outputs can represent all necessary information is unrealistic, especially when the detector is transferred across datasets. In this work, we reason about the graphical model induced by this assumption, and propose to add an auxiliary input to represent missing information such as object relationships. We specifically propose to mine attributes and relationships from the Visual Genome dataset and condition the captioning model on them. Crucially, we propose (and show to be important) the use of a multi-modal pre-trained model (CLIP) to retrieve such contextual descriptions. Further, the object detector outputs are fixed due to a frozen model and hence do not have sufficient richness to allow the captioning model to properly ground them. As a result, we propose to condition both the detector and description outputs on the image, and show qualitatively that this can improve grounding. We validate our method on image captioning, perform thorough analyses of each component and importance of the pre-trained multi-modal model, and demonstrate significant improvements over the current state of the art, specifically +7.5% in CIDEr and +1.3% in BLEU-4 metrics.

By leveraging contrastive learning, clustering, and other pretext tasks, unsupervised methods for learning image representations have reached impressive results on standard benchmarks. The result has been a crowded field -- many methods with substantially different implementations yield results that seem nearly identical on popular benchmarks, such as linear evaluation on ImageNet. However, a single result does not tell the whole story. In this paper, we compare methods using performance-based benchmarks such as linear evaluation, nearest neighbor classification, and clustering for several different datasets, demonstrating the lack of a clear frontrunner within the current state-of-the-art. In contrast to prior work that performs only supervised vs. unsupervised comparison, we compare several different unsupervised methods against each other. To enrich this comparison, we analyze embeddings with measurements such as uniformity, tolerance, and centered kernel alignment (CKA), and propose two new metrics of our own: nearest neighbor graph similarity and linear prediction overlap. We reveal through our analysis that in isolation, single popular methods should not be treated as though they represent the field as a whole, and that future work ought to consider how to leverage the complimentary nature of these methods. We also leverage CKA to provide a framework to robustly quantify augmentation invariance, and provide a reminder that certain types of invariance will be undesirable for downstream tasks.

Weakly-supervised instance segmentation (WSIS) has been considered as a more challenging task than weakly-supervised semantic segmentation (WSSS). Compared to WSSS, WSIS requires instance-wise localization, which is difficult to extract from image-level labels. To tackle the problem, most WSIS approaches use off-the-shelf proposal techniques that require pre-training with instance or object level labels, deviating the fundamental definition of the fully-image-level supervised setting. In this paper, we propose a novel approach including two innovative components. First, we propose a semantic knowledge transfer to obtain pseudo instance labels by transferring the knowledge of WSSS to WSIS while eliminating the need for the off-the-shelf proposals. Second, we propose a self-refinement method to refine the pseudo instance labels in a self-supervised scheme and to use the refined labels for training in an online manner. Here, we discover an erroneous phenomenon, semantic drift, that occurred by the missing instances in pseudo instance labels categorized as background class. This semantic drift occurs confusion between background and instance in training and consequently degrades the segmentation performance. We term this problem as semantic drift problem and show that our proposed self-refinement method eliminates the semantic drift problem. The extensive experiments on PASCAL VOC 2012 and MS COCO demonstrate the effectiveness of our approach, and we achieve a considerable performance without off-the-shelf proposal techniques. The code is available at https://github.com/clovaai/BESTIE.

Recently, a surge of interest in visual transformers is to reduce the computational cost by limiting the calculation of self-attention to a local window. Most current work uses a fixed single-scale window for modeling by default, ignoring the impact of window size on model performance. However, this may limit the modeling potential of these window-based models for multi-scale information. In this paper, we propose a novel method, named Dynamic Window Vision Transformer (DW-ViT). To the best of our knowledge, we are the first to use dynamic multi-scale windows to explore the upper limit of the effect of window settings on model performance. In DW-ViT, multi-scale information is obtained by assigning windows of different sizes to different head groups of window multi-head self-attention. Then, the information is dynamically fused by assigning different weights to the multi-scale window branches. We conducted a detailed performance evaluation on three datasets, ImageNet-1K, ADE20K, and COCO. Compared with related state-of-the-art (SoTA) methods, DW-ViT obtains the best performance. Specifically, compared with the current SoTA Swin Transformers [??], DW-ViT has achieved consistent and substantial improvements on all three datasets with similar parameters and computational costs. In addition, DW-ViT exhibits good scalability and can be easily inserted into any window-based visual transformers.

This paper addresses the problem of vehicle-mounted camera localization by matching a ground-level image with an overhead-view satellite map. Existing methods often treat this problem as cross-view image retrieval, and use learned deep features to match the ground-level query image to a partition (e.g., a small patch) of the satellite map. By these methods, the localization accuracy is limited by the partitioning density of the satellite map (often in the order of tens meters). Departing from the conventional wisdom of image retrieval, this paper presents a novel solution that can achieve highly-accurate localization. The key idea is to formulate the task as pose estimation and solve it by neural-net based optimization. Specifically, we design a two-branch CNN to extract robust features from the ground and satellite images, respectively. To bridge the vast cross-view domain gap, we resort to a Geometry Projection module that projects features from the satellite map to the ground-view, based on a relative camera pose. Aiming to minimize the differences between the projected features and the observed features, we employ a differentiable Levenberg-Marquardt (LM) module to search for the optimal camera pose iteratively. The entire pipeline is differentiable and runs end-to-end. Extensive experiments on standard autonomous vehicle localization datasets have confirmed the superiority of the proposed method. Notably, e.g., starting from a coarse estimate of camera location within a wide region of 40m x 40m, with an 80% likelihood our method quickly reduces the lateral location error to be within 5m on a new KITTI cross-view dataset.

Panoramic images with their 360deg directional view encompass exhaustive information about the surrounding space, providing a rich foundation for scene understanding. To unfold this potential in the form of robust panoramic segmentation models, large quantities of expensive, pixel-wise annotations are crucial for success. Such annotations are available, but predominantly for narrow-angle, pinhole-camera images which, off the shelf, serve as sub-optimal resources for training panoramic models. Distortions and the distinct image-feature distribution in 360deg panoramas impede the transfer from the annotation-rich pinhole domain and therefore come with a big dent in performance. To get around this domain difference and bring together semantic annotations from pinhole- and 360deg surround-visuals, we propose to learn object deformations and panoramic image distortions in the Deformable Patch Embedding (DPE) and Deformable MLP (DMLP) components which blend into our Transformer for PAnoramic Semantic Segmentation (Trans4PASS) model. Finally, we tie together shared semantics in pinhole- and panoramic feature embeddings by generating multi-scale prototype features and aligning them in our Mutual Prototypical Adaptation (MPA) for unsupervised domain adaptation. On the indoor Stanford2D3D dataset, our Trans4PASS with MPA maintains comparable performance to fully-supervised state-of-the-arts, cutting the need for over 1,400 labeled panoramas. On the outdoor DensePASS dataset, we break state-of-the-art by 14.39% mIoU and set the new bar at 56.38%.

Shape matching has been a long-studied problem for the computer graphics and vision community. The objective is to predict a dense correspondence between meshes that have a certain degree of deformation. Existing methods either consider the local description of sampled points or discover correspondences based on global shape information. In this work, we investigate a hierarchical learning design, to which we incorporate local patch-level information and global shape-level structures. This flexible representation enables correspondence prediction and provides rich features for the matching stage. Finally, we propose a novel optimal transport solver by recurrently updating features on non-confident nodes to learn globally consistent correspondences between the shapes. Our results on publicly available datasets suggest robust performance in presence of severe deformations without the need of extensive training or refinement.

Multiple-instance learning (MIL) provides an effective way to tackle the video anomaly detection problem by modeling it as a weakly supervised problem as the labels are usually only available at the video level while missing for frames due to expensive labeling cost. We propose to conduct novel Bayesian non-parametric submodular video partition (BN-SVP) to significantly improve MIL model training that can offer a highly reliable solution for robust anomaly detection in practical settings that include outlier segments or multiple types of abnormal events. BN-SVP essentially performs dynamic non-parametric hierarchical clustering with an enhanced self-transition that groups segments in a video into temporally consistent and semantically coherent hidden states that can be naturally interpreted as scenes. Each segment is assumed to be generated through a non-parametric mixture process that allows variations of segments within the same scenes to accommodate the dynamic and noisy nature of many real-world surveillance videos. The scene and mixture component assignment of BN-SVP also induces a pairwise similarity among segments, resulting in non-parametric construction of a submodular set function. Integrating this function with an MIL loss effectively exposes the model to a diverse set of potentially positive instances to improve its training. A greedy algorithm is developed to optimize the submodular function and support efficient model training. Our theoretical analysis ensures a strong performance guarantee of the proposed algorithm. The effectiveness of the proposed approach is demonstrated over multiple real-world anomaly video datasets with robust detection performance.

Despite the success of deep neural networks, there are still many challenges in deep representation learning due to the data scarcity issues such as data imbalance, unseen distribution, and domain shift. To address the above-mentioned issues, a variety of methods have been devised to explore the sample relationships in a vanilla way (i.e., from the perspectives of either the input or the loss function), failing to explore the internal structure of deep neural networks for learning with sample relationships. Inspired by this, we propose to enable deep neural networks themselves with the ability to learn the sample relationships from each mini-batch. Specifically, we introduce a batch transformer module or BatchFormer, which is then applied into the batch dimension of each mini-batch to implicitly explore sample relationships during training. By doing this, the proposed method enables the collaboration of different samples, e.g., the head-class samples can also contribute to the learning of the tail classes for long-tailed recognition. Furthermore, to mitigate the gap between training and testing, we share the classifier between with or without the BatchFormer during training, which can thus be removed during testing. We perform extensive experiments on over ten datasets and the proposed method achieves significant improvements on different data scarcity applications without any bells and whistles, including the tasks of long-tailed recognition, compositional zero-shot learning, domain generalization, and contrastive learning. Code is made publicly available at https://github.com/zhihou7/BatchFormer.

A recurrent structure is a popular framework choice for the task of video super-resolution. The state-of-the-art method BasicVSR adopts bidirectional propagation with feature alignment to effectively exploit information from the entire input video. In this study, we redesign BasicVSR by proposing second-order grid propagation and flow-guided deformable alignment. We show that by empowering the recurrent framework with enhanced propagation and alignment, one can exploit spatiotemporal information across misaligned video frames more effectively. The new components lead to an improved performance under a similar computational constraint. In particular, our model BasicVSR++ surpasses BasicVSR by a significant 0.82 dB in PSNR with similar number of parameters. BasicVSR++ is generalizable to other video restoration tasks, and obtains three champions and one first runner-up in NTIRE 2021 video restoration challenge.

Instance-level feature matching is significantly important to the success of modern one-shot object detectors. Recently, the methods based on the metric-learning paradigm have achieved an impressive process. Most of these works only measure the relations between query and target objects on a single level, resulting in suboptimal performance overall. In this paper, we introduce the balanced and hierarchical learning for our detector. The contributions are two-fold: firstly, a novel Instance-level Hierarchical Relation (IHR) module is proposed to encode the contrastive-level, salient-level, and attention-level relations simultaneously to enhance the query-relevant similarity representation. Secondly, we notice that the batch training of the IHR module is substantially hindered by the positive-negative sample imbalance in the one-shot scenario. We then introduce a simple but effective Ratio-Preserving Loss (RPL) to protect the learning of rare positive samples and suppress the effects of negative samples. Our loss can adjust the weight for each sample adaptively, ensuring the desired positive-negative ratio consistency and boosting query-related IHR learning. Extensive experiments show that our method outperforms the state-of-the-art method by 1.6% and 1.3% on PASCAL VOC and MS COCO datasets for unseen classes, respectively. The code will be available at https://github.com/hero-y/BHRL.

Real-world data typically follow a long-tailed distribution, where a few majority categories occupy most of the data while most minority categories contain a limited number of samples. Classification models minimizing cross-entropy struggle to represent and classify the tail classes. Although the problem of learning unbiased classifiers has been well studied, methods for representing imbalanced data are under-explored. In this paper, we focus on representation learning for imbalanced data. Recently, supervised contrastive learning has shown promising performance on balanced data recently. However, through our theoretical analysis, we find that for long-tailed data, it fails to form a regular simplex which is an ideal geometric configuration for representation learning. To correct the optimization behavior of SCL and further improve the performance of long-tailed visual recognition, we propose a novel loss for balanced contrastive learning (BCL). Compared with SCL, we have two improvements in BCL: class-averaging, which balances the gradient contribution of negative classes; class-complement, which allows all classes to appear in every mini-batch. The proposed balanced contrastive learning (BCL) method satisfies the condition of forming a regular simplex and assists the optimization of cross-entropy. Equipped with BCL, the proposed two-branch framework can obtain a stronger feature representation and achieve competitive performance on long-tailed benchmark datasets such as CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist2018.

Weakly supervised object localization (WSOL) aims to localize objects using only image-level labels. Recently a new paradigm has emerged by generating a foreground prediction map (FPM) to achieve localization task. Existing FPM-based methods use cross-entropy (CE) to evaluate the foreground prediction map and to guide the learning of generator. We argue for using activation value to achieve more efficient learning. It is based on the experimental observation that, for a trained network, CE converges to zero when the foreground mask covers only part of the object region. While activation value increases until the mask expands to the object boundary, which indicates that more object areas can be learned by using activation value. In this paper, we propose a Background Activation Suppression (BAS) method. Specifically, an Activation Map Constraint module (AMC) is designed to facilitate the learning of generator by suppressing the background activation value. Meanwhile, by using the foreground region guidance and the area constraint, BAS can learn the whole region of the object. In the inference phase, we consider the prediction maps of different categories together to obtain the final localization results. Extensive experiments show that BAS achieves significant and consistent improvement over the baseline methods on the CUB-200-2011 and ILSVRC datasets. Code and models are available at https://github.com/wpy1999/BAS.

In this paper, we propose a weakly-supervised approach for 3D object detection, which makes it possible to train a strong 3D detector with position-level annotations (i.e. annotations of object centers). In order to remedy the information loss from box annotations to centers, our method, namely Back to Reality (BR), makes use of synthetic 3D shapes to convert the weak labels into fully-annotated virtual scenes as stronger supervision, and in turn utilizes the perfect virtual labels to complement and refine the real labels. Specifically, we first assemble 3D shapes into physically reasonable virtual scenes according to the coarse scene layout extracted from position-level annotations. Then we go back to reality by applying a virtual-to-real domain adaptation method, which refine the weak labels and additionally supervise the training of detector with the virtual scenes. With less than 5% of the labeling labor, we achieve comparable detection performance with some popular fully-supervised approaches on the widely used ScanNet dataset.

Differentiable Architecture Search (DARTS) has received massive attention in recent years, mainly because it significantly reduces the computational cost through weight sharing and continuous relaxation. However, more recent works find that existing differentiable NAS techniques struggle to outperform naive baselines, yielding deteriorative architectures as the search proceeds. Rather than directly optimizing the architecture parameters, this paper formulates the neural architecture search as a distribution learning problem through relaxing the architecture weights into Gaussian distributions. By leveraging the natural-gradient variational inference (NGVI), the architecture distribution can be easily optimized based on existing codebases without incurring more memory and computational consumption. We demonstrate how the differentiable NAS benefits from Bayesian principles, enhancing exploration and improving stability. The experimental results on NAS benchmark datasets confirm the significant improvements the proposed framework can make. In addition, instead of simply applying the argmax on the learned parameters, we further leverage the recently-proposed training-free proxies in NAS to select the optimal architecture from a group architectures drawn from the optimized distribution, where we achieve state-of-the-art results on the NAS-Bench-201 and NAS-Bench-1shot1 benchmarks. Our best architecture in the DARTS search space also obtains competitive test errors with 2.37%, 15.72%, and 24.2% on CIFAR-10, CIFAR-100, and ImageNet, respectively.