ICCV 2021 论文列表

Developing deep neural networks to generate 3D scenes is a fundamental problem in neural synthesis with immediate applications in architectural CAD, computer graphics, as well as in generating virtual robot training environments. This task is challenging because 3D scenes exhibit diverse patterns, ranging from continuous ones, such as object sizes and the relative poses between pairs of shapes, to discrete patterns, such as occurrence and co-occurrence of objects with symmetrical relationships. This paper introduces a novel neural scene synthesis approach that can capture diverse feature patterns of 3D scenes. Our method combines the strength of both neural network-based and conventional scene synthesis approaches. We use the parametric prior distributions learned from training data, which provide uncertainties of object attributes and relative attributes, to regularize the outputs of feed-forward neural models. Moreover, instead of merely predicting a scene layout, our approach predicts an over-complete set of attributes. This methodology allows us to utilize the underlying consistency constraints among the predicted attributes to prune infeasible predictions. Experimental results show that our approach outperforms existing methods considerably. The generated 3D scenes interpolate the training data faithfully while preserving both continuous and discrete feature patterns.

Salient object detection identifies objects in an image that grab visual attention. Although contextual features are considered in recent literature, they often fail in real-world complex scenarios. We observe that this is mainly due to two issues: First, most existing datasets consist of simple foregrounds and backgrounds that hardly represent real-life scenarios. Second, current methods only learn contextual features of salient objects, which are insufficient to model high-level semantics for saliency reasoning in complex scenes. To address these problems, we first construct a new large-scale dataset with complex scenes in this paper. We then propose a context-aware learning approach to explicitly exploit the semantic scene contexts. Specifically, two modules are proposed to achieve the goal: 1) a Semantic Scene Context Refinement module to enhance contextual features learned from salient objects with scene context, and 2) a Contextual Instance Transformer to learn contextual relations between objects and scene context. To our knowledge, such high-level semantic contextual information of image scenes is under-explored for saliency detection in the literature. Extensive experiments demonstrate that the proposed approach outperforms state-of-the-art techniques in complex scenarios for saliency detection, and transfers well to other existing datasets. The code and dataset are available at https://github.com/SirisAvishek/Scene_Context_Aware_Saliency.

Image translation methods typically aim to manipulate a set of labeled attributes (given as supervision at training time e.g. domain label) while leaving the unlabeled attributes intact. Current methods achieve either: (i) disentanglement, which exhibits low visual fidelity and can only be satisfied where the attributes are perfectly uncorrelated. (ii) visually-plausible translations, which are clearly not disentangled. In this work, we propose OverLORD, a single framework for disentangling labeled and unlabeled attributes as well as synthesizing high-fidelity images, which is composed of two stages; (i) Disentanglement: Learning disentangled representations with latent optimization. Differently from previous approaches, we do not rely on adversarial training or any architectural biases. (ii) Synthesis: Training feed-forward encoders for inferring the learned attributes and tuning the generator in an adversarial manner to increase the perceptual quality. When the labeled and unlabeled attributes are correlated, we model an additional representation that accounts for the correlated attributes and improves disentanglement. We highlight that our flexible framework covers multiple settings as disentangling labeled attributes, pose and appearance, localized concepts, and shape and texture. We present significantly better disentanglement with higher translation quality and greater output diversity than state-of-the-art methods.

Manually annotating object segmentation masks is very time-consuming. While interactive segmentation methods offer a more efficient alternative, they become unaffordable at a large scale because the cost grows linearly with the number of annotated masks. In this paper, we propose a highly efficient annotation scheme for building large datasets with object segmentation masks. At a large scale, images contain many object instances with similar appearance. We exploit these similarities by using hierarchical clustering on mask predictions made by a segmentation model. We propose a scheme that efficiently searches through the hierarchy of clusters and selects which clusters to annotate. Humans manually verify only a few masks per cluster, and the labels are propagated to the whole cluster. Through a large-scale experiment to populate 1M unlabeled images with object segmentation masks for 80 object classes, we show that (1) we obtain 1M object segmentation masks with an total annotation time of only 290 hours; (2) we reduce annotation time by 76x compared to manual annotation; (3) the segmentation quality of our masks is on par with those from manually annotated datasets. Code, data, and models are available online.

The state-of-the-art object detection and image classification methods can perform impressively on more than 9k and 10k classes respectively. In contrast, the number of classes in semantic segmentation datasets is relatively limited. This is not surprising when the restrictions caused by the lack of labelled data and high computation demand for segmentation are considered. In this paper, we propose a novel training methodology to train and scale the existing semantic segmentation models for a large number of semantic classes without increasing the memory overhead. In our approach, we reduce the space complexity of the segmentation model's output from O(C) to O(1), propose an approximation method for ground-truth class probability, and use it to compute cross-entropy loss. The proposed approach is general and can be adopted by any state-of-the-art segmentation model to gracefully scale it for any number of semantic classes with only one GPU. Our approach achieves similar, and in some cases even better mIoU for Cityscapes, Pascal VOC and ADE20k dataset when adopted to DeeplabV3+ model with different backbones. We demonstrate a clear benefit of our approach on a dataset with 1284 classes, bootstrapped from LVIS and COCO annotations, with almost three times better mIoU when compared to DeeplabV3+. Code is available at: https://github.com/shipra25jain/ESSNet.

The recently proposed Visual image Transformers (ViT) with pure attention have achieved promising performance on image recognition tasks, such as image classification. However, the routine of the current ViT model is to maintain a full-length patch sequence during inference, which is redundant and lacks hierarchical representation. To this end, we propose a Hierarchical Visual Transformer (HVT) which progressively pools visual tokens to shrink the sequence length and hence reduces the computational cost, analogous to the feature maps downsampling in Convolutional Neural Networks (CNNs). It brings a great benefit that we can increase the model capacity by scaling dimensions of depth/width/resolution/patch size without introducing extra computational complexity due to the reduced sequence length. Moreover, we empirically find that the average pooled visual tokens contain more discriminative information than the single class token. To demonstrate the improved scalability of our HVT, we conduct extensive experiments on the image classification task. With comparable FLOPs, our HVT outperforms the competitive baselines on ImageNet and CIFAR-100 datasets. Code is available at https://github.com/MonashAI/HVT.

We present a novel method for synthesizing both temporally and geometrically consistent street-view panoramic video from a single satellite image and camera trajectory. Existing cross-view synthesis approaches focus on images, while video synthesis in such a case has not yet received enough attention. For geometrical and temporal consistency, our approach explicitly creates a 3D point cloud representation of the scene and maintains dense 3D-2D correspondences across frames that reflect the geometric scene configuration inferred from the satellite view. As for synthesis in the 3D space, we implement a cascaded network architecture with two hourglass modules to generate point-wise coarse and fine features from semantics and per-class latent vectors, followed by projection to frames and an upsampling module to obtain the final realistic video. By leveraging computed correspondences, the produced street-view video frames adhere to the 3D geometric scene structure and maintain temporal consistency. Qualitative and quantitative experiments demonstrate superior results compared to other state-of-the-art synthesis approaches that either lack temporal consistency or realistic appearance. To the best of our knowledge, our work is the first one to synthesize cross-view images to videos.

In this paper, by modeling the point cloud registration task as a Markov decision process, we propose an end-to-end deep model embedded with the cross-entropy method (CEM) for unsupervised 3D registration. Our model consists of a sampling network module and a differentiable CEM module. In our sampling network module, given a pair of point clouds, the sampling network learns a prior sampling distribution over the transformation space. The learned sampling distribution can be used as a "good"" initialization of the differentiable CEM module. In our differentiable CEM module, we first propose a maximum consensus criterion based alignment metric as the reward function for the point cloud registration task. Based on the reward function, for each state, we then construct a fused score function to evaluate the sampled transformations, where we weight the current and future rewards of the transformations. Particularly, the future rewards of the sampled transforms are obtained by performing the iterative closest point (ICP) algorithm on the transformed state. By selecting the top-k transformations with the highest scores, we iteratively update the sampling distribution. Furthermore, in order to make the CEM differentiable, we use the sparsemax function to replace the hard top-k selection. Finally, we formulate a Geman-McClure estimator based loss to train our end-to-end registration model. Extensive experimental results demonstrate the good registration performance of our method on benchmark datasets.

Adversarial robustness of deep models is pivotal in ensuring safe deployment in real world settings, but most modern defenses have narrow scope and expensive costs. In this paper, we propose a self-supervised method to detect adversarial attacks and classify them to their respective threat models, based on a linear model operating on the embeddings from a pre-trained self-supervised encoder. We use a SimCLR encoder in our experiments, since we show the SimCLR embedding distance is a good proxy for human perceptibility, enabling it to encapsulate many threat models at once. We call our method SimCat since it uses SimCLR encoder to catch and categorize various types of adversarial attacks, including L_p and non-L_p evasion attacks, as well as data poisonings. The simple nature of a linear classifier makes our method efficient in both time and sample complexity. For example, on SVHN, using only five pairs of clean and adversarial examples computed with a PGD-L_inf attack, SimCat's detection accuracy is over 85%. Moreover, on ImageNet, using only 25 examples from each threat model, SimCat can classify eight different attack types such as PGD-L_2, PGD-L_inf, CW-L_2, PPGD, LPA, StAdv, ReColor, and JPEG-L_inf, with over 40% accuracy. On STL10 data, we apply SimCat as a defense against poisoning attacks, such as BP, CP, FC, CLBD, HTBD, halving the success rate while using only twenty total poisons for training. We find that the detectors generalize well to unseen threat models. Lastly, we investigate the performance of our detection method under adaptive attacks and further boost its robustness against such attacks via adversarial training.

Instance segmentation can detect where the objects are in an image, but hard to understand the relationship between them. We pay attention to a typical relationship, relative saliency. A closely related task, salient object detection, predicts a binary map highlighting a visually salient region while hard to distinguish multiple objects. Directly combining two tasks by post-processing also leads to poor performance. There is a lack of research on relative saliency at present, limiting the practical applications such as content-aware image cropping, video summary, and image labeling. In this paper, we study the Salient Object Ranking (SOR) task, which manages to assign a ranking order of each detected object according to its visual saliency. We propose the first end-to-end framework of the SOR task and solve it in a multi-task learning fashion. The framework handles instance segmentation and salient object ranking simultaneously. In this framework, the SOR branch is independent and flexible to cooperate with different detection methods, so that easy to use as a plugin. We also introduce a Position-Preserved Attention (PPA) module tailored for the SOR branch. It consists of the position embedding stage and feature interaction stage. Considering the importance of position in saliency comparison, we preserve absolute coordinates of objects in ROI pooling operation and then fuse positional information with semantic features in the first stage. In the feature interaction stage, we apply the attention mechanism to obtain proposals' contextualized representations to predict their relative ranking orders. Extensive experiments have been conducted on the ASR dataset. Without bells and whistles, our proposed method outperforms the former state-of-the-art method significantly. The code will be released publicly available on https://github.com/EricFH/SOR.

Most existing trackers based on deep learning perform tracking in a holistic strategy, which aims to learn deep representations of the whole target for localizing the target. It is arduous for such methods to track targets with various appearance variations. To address this limitation, another type of methods adopts a part-based tracking strategy which divides the target into equal patches and tracks all these patches in parallel. The target state is inferred by summarizing the tracking results of these patches. A potential limitation of such trackers is that not all patches are equally informative for tracking. Some patches that are not discriminative may have adverse effects. In this paper, we propose to track the salient local parts of the target that are discriminative for tracking. In particular, we propose a fine-grained saliency mining module to capture the local saliencies. Further, we design a saliency-association modeling module to associate the captured saliencies together to learn effective correlation representations between the exemplar and the search image for state estimation. Extensive experiments on five diverse datasets demonstrate that the proposed method performs favorably against state-of-the-art trackers.

Motion prediction of vehicles is critical but challenging due to the uncertainties in complex environments and the limited visibility caused by occlusions and limited sensor ranges. In this paper, we study a new task, safety-aware motion prediction with unseen vehicles for autonomous driving. Unlike the existing trajectory prediction task for seen vehicles, we aim at predicting an occupancy map that indicates the earliest time when each location can be occupied by either seen and unseen vehicles. The ability to predict unseen vehicles is critical for safety in autonomous driving. To tackle this challenging task, we propose a safety-aware deep learning model with three new loss functions to predict the earliest occupancy map. Experiments on the large-scale autonomous driving nuScenes dataset show that our proposed model significantly outperforms the state-of-the-art baselines on the safety-aware motion prediction task. To the best of our knowledge, our approach is the first one that can predict the existence of unseen vehicles in most cases.

Most monocular depth sensing methods use conventionally captured images that are created without considering scene content. In contrast, animal eyes have fast mechanical motions, called saccades, that control how the scene is imaged by the fovea, where resolution is highest. In this paper, we present the SaccadeCam framework for adaptively distributing resolution onto regions of interest in the scene. Our algorithm for adaptive resolution is a self-supervised network and we demonstrate results for end-to-end learning for monocular depth estimation. We also show preliminary results with a real SaccadeCam hardware prototype.

The vast majority of modern consumer-grade cameras employ a rolling shutter mechanism, leading to image distortions if the camera moves during image acquisition. In this paper, we present a novel deep network to solve the generic rolling shutter correction problem with two consecutive frames. Our pipeline is symmetrically designed to predict the global shutter image corresponding to the intermediate time of these two frames, which is difficult for existing methods because it corresponds to a camera pose that differs most from the two frames. First, two time-symmetric dense undistortion flows are estimated by using well-established principles: pyramidal construction, warping, and cost volume processing. Then, both rolling shutter images are warped into a common global shutter one in the feature space, respectively. Finally, a symmetric consistency constraint is constructed in the image decoder to effectively aggregate the contextual cues of two rolling shutter images, thereby recovering the high-quality global shutter image. Extensive experiments with both synthetic and real data from public benchmarks demonstrate the superiority of our proposed approach over the state-of-the-art methods.

Spatio-temporal video grounding (STVG) aims to localize a spatio-temporal tube of a target object in an untrimmed video based on a query sentence. In this work, we propose a one-stage visual-linguistic transformer based framework called STVGBert for the STVG task, which can simultaneously localize the target object in both spatial and temporal domains. Specifically, without resorting to pre-generated object proposals, our STVGBert directly takes a video and a query sentence as the input, and then produces the cross-modal features by using the newly introduced cross-modal feature learning module ST-ViLBert. Based on the cross-modal features, our method then generates bounding boxes and predicts the starting and ending frames to produce the predicted object tube. To the best of our knowledge, our STVGBert is the first one-stage method, which can handle the STVG task without relying on any pre-trained object detectors. Comprehensive experiments demonstrate our newly proposed framework outperforms the state-of-the-art multi-stage methods on two benchmark datasets Vid-STG and HC-STVG.

We propose replacing scene text in videos using deep style transfer and learned photometric transformations. Building on recent progress on still image text replacement, we present extensions that alter text while preserving the appearance and motion characteristics of the original video. Compared to the problem of still image text replacement, our method addresses additional challenges introduced by video, namely effects induced by changing lighting, motion blur, diverse variations in camera-object pose over time, and preservation of temporal consistency. We parse the problem into three steps. First, the text in all frames is normalized to a frontal pose using a spatio-temporal transformer network. Second, the text is replaced in a single reference frame using a state-of-art still-image text replacement method. Finally, the new text is transferred from the reference to remaining frames using a novel learned image transformation network that captures lighting and blur effects in a temporally consistent manner. Results on synthetic and challenging real videos show realistic text transfer, competitive quantitative and qualitative performance, and superior inference speed relative to alternatives. We introduce new synthetic and real-world datasets with paired text objects. To the best of our knowledge this is the first attempt at deep video text replacement.

Geometry-aware modules are widely applied in recent deep learning architectures for scene representation and rendering. However, these modules require intrinsic camera information that might not be obtained accurately. In this paper, we propose a Spatial Transformation Routing (STR) mechanism to model the spatial properties without applying any geometric prior. The STR mechanism treats the spatial transformation as the message passing process, and the relation between the view poses and the routing weights is modeled by an end-to-end trainable neural network. Besides, an Occupancy Concept Mapping (OCM) framework is proposed to provide explainable rationals for scene-fusion processes. We conducted experiments on several datasets and show that the proposed STR mechanism improves the performance of the Generative Query Network (GQN). The visualization results reveal that the routing process can pass the observed information from one location of some view to the associated location in the other view, which demonstrates the advantage of the proposed model in terms of spatial cognition.

Multi-source Domain Adaptation (MSDA) is more practical but challenging than the conventional unsupervised domain adaptation due to the involvement of diverse multiple data sources. Two fundamental challenges of MSDA are: (i) how to deal with the diversity in the multiple source domains and (ii) how to cope with the data shift between the target domain and the source domains. In this paper, to address the first challenge, we propose a theoretical-guaranteed approach to combine domain experts locally trained on its own source domain to achieve a combined multi-source teacher that globally predicts well on the mixture of source domains. To address the second challenge, we propose to bridge the gap between the target domain and the mixture of source domains in the latent space via a generator or feature extractor. Together with bridging the gap in the latent space, we train a student to mimic the predictions of the teacher expert on both source and target examples. In addition, our approach is guaranteed with rigorous theory offered insightful justifications of how each component influences the transferring performance. Extensive experiments conducted on three benchmark datasets show that our proposed method achieves state-of-the-art performances to the best of our knowledge.

Image and video enhancement such as color constancy, low light enhancement, and tone mapping on smartphones is challenging because high-quality images should be achieved efficiently with a limited resource budget. Unlike prior works that either used very deep CNNs or large Transformer models, we propose a \underline s eman\underline t ic-\underline a wa\underline r e lightweight Transformer, termed STAR, for real-time image enhancement. STAR is formulated to capture long-range dependencies between image patches, which naturally and implicitly captures the semantic relationships of different regions in an image. STAR is a general architecture that can be easily adapted to different image enhancement tasks. Extensive experiments show that STAR can effectively boost the quality and efficiency of many tasks such as illumination enhancement, auto white balance, and photo retouching, which are indispensable components for image processing on smartphones. For example, STAR reduces model complexity and improves image quality compared to the recent state-of-the-art [??] on the MIT-Adobe FiveK dataset [??] (i.e., 1.8dB PSNR improvements with 25% parameters and 13% float operations.)