CVPR 2021 论文列表

We address the task of converting a floorplan and a set of associated photos of a residence into a textured 3D mesh model, a task which we call Plan2Scene. Our system 1) lifts a floorplan image to a 3D mesh model; 2) synthesizes surface textures based on the input photos; and 3) infers textures for unobserved surfaces using a graph neural network architecture. To train and evaluate our system we create indoor surface texture datasets, and augment a dataset of floorplans and photos from prior work with rectified surface crops and additional annotations. Our approach handles the challenge of producing tileable textures for dominant surfaces such as floors, walls, and ceilings from a sparse set of unaligned photos that only partially cover the residence. Qualitative and quantitative evaluations show that our system produces realistic 3D interior models, outperforming baseline approaches on a suite of texture quality metrics and as measured by a holistic user study.

The inability of state-of-the-art semantic segmentation methods to detect anomaly instances hinders them from being deployed in safety-critical and complex applications, such as autonomous driving. Recent approaches have focused on either leveraging segmentation uncertainty to identify anomalous areas or re-synthesizing the image from the semantic label map to find dissimilarities with the input image. In this work, we demonstrate that these two methodologies contain complementary information and can be combined to produce robust predictions for anomaly segmentation. We present a pixel-wise anomaly detection framework that uses uncertainty maps to improve over existing re-synthesis methods in finding dissimilarities between the input and generated images. Our approach works as a general framework around already trained segmentation networks, which ensures anomaly detection without compromising segmentation accuracy, while significantly outperforming all similar methods. Top-2 performance across a range of different anomaly datasets shows the robustness of our approach to handling different anomaly instances.

Acquisition and rendering of photo-realistic human heads is a highly challenging research problem of particular importance for virtual telepresence. Currently, the highest quality is achieved by volumetric approaches trained in a person-specific manner on multi-view data. These models better represent fine structure, such as hair, compared to simpler mesh-based models. Volumetric models typically employ a global code to represent facial expressions, such that they can be driven by a small set of animation parameters. While such architectures achieve impressive rendering quality, they can not easily be extended to the multi-identity setting. In this paper, we devise a novel approach for predicting volumetric avatars of the human head given just a small number of inputs. We enable generalization across identities by a novel parameterization that combines neural radiance fields with local, pixel-aligned features extracted directly from the inputs, thus side-stepping the need for very deep or complex networks. Our approach is trained in an end-to-end manner solely based on a photometric re-rendering loss without requiring explicit 3D supervision. We demonstrate that our approach outperforms the existing state of the art in terms of quality and is able to generate faithful facial expressions in a multi-identity setting.

Telecommunication with photorealistic avatars in virtual or augmented reality is a promising path for achieving authentic face-to-face communication in 3D over remote physical distances. In this work, we present the Pixel Codec Avatars (PiCA): a deep generative model of 3D human faces that achieves state of the art reconstruction performance while being computationally efficient and adaptive to the rendering conditions during execution. Our model combines two core ideas: (1) a fully convolutional architecture for decoding spatially varying features, and (2) a rendering-adaptive per-pixel decoder. Both techniques are integrated via a dense surface representation that is learned in a weakly-supervised manner from low-topology mesh tracking over training images. We demonstrate that PiCA improves reconstruction over existing techniques across testing expressions and views on persons of different gender and skin tone. Importantly, we show that the PiCA model is much smaller than the state-of-art baseline model, and makes multi-person telecommunicaiton possible: on a single Oculus Quest 2 mobile VR headset, 5 avatars are rendered in realtime in the same scene.

Unsupervised domain adaptation is a promising technique for semantic segmentation and other computer vision tasks for which large-scale data annotation is costly and time-consuming. In semantic segmentation particularly, it is attractive to train models on annotated images from a simulated (source) domain and deploy them on real (target) domains. In this work, we present a novel framework for unsupervised domain adaptation based on the notion of target-domain consistency training. Intuitively, our work is based on the insight that in order to perform well on the target domain, a model's output should be consistent with respect to small perturbations of inputs in the target domain. Specifically, we introduce a new loss term to enforce pixelwise consistency between the model's predictions on a target image and perturbed version of the same image. In comparison to popular adversarial adaptation methods, our approach is simpler, easier to implement, and more memory-efficient during training. Experiments and ablation studies demonstrate that our simple approach achieves remarkably strong results on two challenging synthetic-to-real benchmarks, GTA5-to-Cityscapes and SYNTHIA-to-Cityscapes.

We present Picasso, a CUDA-based library comprising novel modules for deep learning over complex real-world 3D meshes. Hierarchical neural architectures have proved effective in multi-scale feature extraction which signifies the need for fast mesh decimation. However, existing methods rely on CPU-based implementations to obtain multi-resolution meshes. We design GPU-accelerated mesh decimation to facilitate network resolution reduction efficiently on-the-fly. Pooling and unpooling modules are defined on the vertex clusters gathered during decimation. For feature learning over meshes, Picasso contains three types of novel convolutions namely, facet2vertex, vertex2facet, and facet2facet convolution. Hence, it treats a mesh as a geometric structure comprising vertices and facets, rather than a spacial graph with edges as previous methods do. Picasso also incorporates a fuzzy mechanism in its filters for robustness to mesh sampling (vertex density). It exploits Gaussian mixtures to define fuzzy coefficients for the facet2vertex convolution, and barycentric interpolation to define the coefficients for the remaining two convolutions. In this release, we demonstrate the effectivenss of the proposed modules with competitive segmentation results on S3DIS. The library will be made public through github.

We present a new framework for semantic segmentation without annotations via clustering. Off-the-shelf clustering methods are limited to curated, single-label, and object-centric images yet real-world data are dominantly uncurated, multi-label, and scene-centric. We extend clustering from images to pixels and assign separate cluster membership to different instances within each image. However, solely relying on pixel-wise feature similarity fails to learn high-level semantic concepts and overfits to low-level visual cues. We propose a method to incorporate geometric consistency as an inductive bias to learn invariance and equivariance for photometric and geometric variations. With our novel learning objective, our framework can learn high-level semantic concepts. Our method, PiCIE (Pixel-level feature Clustering using Invariance and Equivariance), is the first method capable of segmenting both things and stuff categories without any hyperparameter tuning or task-specific pre-processing. Our method largely outperforms existing baselines on COCO and Cityscapes with +17.5 Acc. and +4.5 mIoU. We show that PiCIE gives a better initialization for standard supervised training. The code is available at https:// github.com/janghyuncho/PiCIE.

We have witnessed rapid progress on 3D-aware image synthesis, leveraging recent advances in generative visual models and neural rendering. Existing approaches however fall short in two ways: first, they may lack an underlying 3D representation or rely on view-inconsistent rendering, hence synthesizing images that are not multi-view consistent; second, they often depend upon representation network architectures that are not expressive enough, and their results thus lack in image quality. We propose a novel generative model, named Periodic Implicit Generative Adversarial Networks (p-GAN or pi-GAN), for high-quality 3D-aware image synthesis. p-GAN leverages neural representations with periodic activation functions and volumetric rendering to represent scenes as view-consistent radiance fields. The proposed approach obtains state-of-the-art results for 3D-aware image synthesis with multiple real and synthetic datasets.

Phase retrieval from intensity-only measurements plays a central role in many real-world imaging tasks. In recent years, deep neural networks based methods emerge and show promising performance for phase retrieval. However, their interpretability and generalization still remain a major challenge. In this paper, we propose to combine the advantages of both model-based alternative projection method and deep neural network for phase retrieval, so as to achieve network interpretability and inference effectiveness simultaneously. Specifically, we unfold the iterative process of the alternative projection phase retrieval into a feed-forward neural network, whose layers mimic the processing flow. The physical model of the imaging process is then naturally embedded into the neural network structure. Moreover, a complex-valued U-Net is proposed for defining image priori for forward and backward projection in dual planes. Finally, we designate physics-based formulation as an untrained deep neural network, whose weights are enforced to fit to the given intensity measurements. In summary, our scheme for phase retrieval is effective, interpretable, physics-based and unsupervised. Experimental results demonstrate that our method achieves superior performance compared with the state-of-the-arts in a practical phase retrieval application---lensless microscopy imaging.

Shapes are often designed to satisfy structural properties and serve a particular functionality in the physical world. Unfortunately, most existing generative models focus primarily on the geometric or visual plausibility, ignoring the physical or structural constraints. To remedy this, we present a novel method aimed to endow deep generative models with physical reasoning. In particular, we introduce a loss and a learning framework that promote two key characteristics of the generated shapes: their connectivity and physical stability. The former ensures that each generated shape consists of a single connected component, while the latter promotes the stability of that shape when subjected to gravity. Our proposed physical losses are fully differentiable and we demonstrate their use in end-to-end learning. Crucially we demonstrate that such physical objectives can be achieved without sacrificing the expressive power of the model and variability of the generated results. We demonstrate through extensive comparisons with the state-of-the-art deep generative models, the utility and efficiency of our proposed approach, while avoiding the potentially costly differentiable physical simulation at training time.

We present an end-to-end inverse rendering pipeline that includes a fully differentiable renderer, and can reconstruct geometry, materials, and illumination from scratch from a set of images. Our rendering framework represents specular BRDFs and environmental illumination using mixtures of spherical Gaussians, and represents geometry as a signed distance function parameterized as a Multi-Layer Perceptron. The use of spherical Gaussians allows us to efficiently solve for approximate light transport, and our method works on scenes with challenging non-Lambertian reflectance captured under natural, static illumination. We demonstrate, with both synthetic and real data, that our reconstruction not only can render novel viewpoints, but also enables physics-based appearance editing of materials and illumination.

Vehicle re-identification (re-ID) is of great significance to urban operation, management, security and has gained more attention in recent years. However, two critical challenges in vehicle re-ID have primarily been underestimated, i.e., 1): how to make full use of raw data, and 2): how to learn a robust re-ID model with noisy data. In this paper, we first create a video vehicle re-ID evaluation benchmark called VVeRI-901 and verify the performance of video-based re-ID is far better than static image-based one. Then we propose a new Pompeiu-hausdorff distance (PhD) learning method for video-to-video matching. It can alleviate the data noise problem caused by the occlusion in videos and thus improve re-ID performance significantly. Extensive empirical results on video-based vehicle and person re-ID datasets, i.e., VVeRI-901, MARS and PRID2011, demonstrate the superiority of the proposed method. The source code of our proposed method is available at https://github.com/emdata-ailab/PhD-Learning.

The multimedia community has recently seen a tremendous surge of interest in the fashion recommendation problem. A lot of efforts have been made to model the compatibility between fashion items. Some have also studied users' personal preferences for the outfits. There is, however, another difficulty in the task that hasn't been dealt with carefully by previous work. Users that are new to the system usually only have several (less than 5) outfits available for learning. With such a limited number of training examples, it is challenging to model the user's preferences reliably. In this work, we propose a new solution for personalized outfit recommendation that is capable of handling this case. We use a stacked self-attention mechanism to model the high-order interactions among the items. We then embed the items in an outfit into a single compact representation within the outfit space. To accommodate the variety of users' preferences, we characterize each user with a set of anchors, i.e. a group of learnable latent vectors in the outfit space that are the representatives of the outfits the user likes. We also learn a set of general anchors to model the general preference shared by all users. Based on this representation of the outfits and the users, we propose a simple but effective strategy for the new user profiling tasks. Extensive experiments on large scale real-world datasets demonstrate the performance of our proposed method.

Recently, person re-identification (ReID) has vastly benefited from the surging waves of data-driven methods. However, these methods are still not reliable enough for real-world deployments, due to the insufficient generalization capability of the models learned on existing benchmarks that have limitations in multiple aspects, including limited data scale, capture condition variations, and appearance diversities. To this end, we collect a new dataset named Person30K with the following distinct features: 1) a very large scale containing 1.38 million images of 30K identities, 2) a large capture system containing 6,497 cameras deployed at 89 different sites, 3) abundant sample diversities including varied backgrounds and diverse person poses. Furthermore, we propose a domain generalization ReID method, dual-meta generalization network (DMG-Net), to exploit the merits of meta-learning in both the training procedure and the metric space learning. Concretely, we design a "learning then generalization evaluation" meta-training procedure and a meta-discrimination loss to enhance model generalization and discrimination capabilities. Comprehensive experiments validate the effectiveness of our DMG-Net. (Dataset and code will be released.)

Recently, some methods have focused on learning local relation among parts of pedestrian images for person re-identification (Re-ID), as it offers powerful representation capabilities. However, they only provide the intra-local relation among parts within single pedestrian image and ignore the inter-local relation among parts from different images, which results in incomplete local relation information. In this paper, we propose a novel deep graph model named Heterogeneous Local Graph Attention Networks (HLGAT) to model the inter-local relation and the intra-local relation in the completed local graph, simultaneously. Specifically, we first construct the completed local graph using local features, and we resort to the attention mechanism to aggregate the local features in the learning process of inter-local relation and intra-local relation so as to emphasize the importance of different local features. As for the inter-local relation, we propose the attention regularization loss to constrain the attention weights based on the identities of local features in order to describe the inter-local relation accurately. As for the intra-local relation, we propose to inject the contextual information into the attention weights to consider structure information. Extensive experiments on Market-1501, CUHK03, DukeMTMC-reID and MSMT17 demonstrate that the proposed HLGAT outperforms the state-of-the-art methods.

Recent work has shown that convolutional neural network classifiers overly rely on texture at the expense of shape cues. We make a similar but different distinction between shape and local image cues, on the one hand, and global image statistics, on the other. Our method, called Permuted Adaptive Instance Normalization (pAdaIN), reduces the representation of global statistics in the hidden layers of image classifiers. pAdaIN samples a random permutation p that rearranges the samples in a given batch. Adaptive Instance Normalization (AdaIN) is then applied between the activations of each (non-permuted) sample i and the corresponding activations of the sample p(i), thus swapping statistics between the samples of the batch. Since the global image statistics are distorted, this swapping procedure causes the network to rely on cues, such as shape or texture. By choosing the random permutation with probability p and the identity permutation otherwise, one can control the effect's strength. With the correct choice of p, fixed apriori for all experiments and selected without considering test data, our method consistently outperforms baselines in multiple settings. In image classification, our method improves on both CIFAR100 and ImageNet using multiple architectures. In the setting of robustness, our method improves on both ImageNet-C and Cifar-100-C for multiple architectures. In the setting of domain adaptation and domain generalization, our method achieves state of the art results on the transfer learning task from GTAV to Cityscapes and on the PACS benchmark.

Compressing large neural networks is an important step for their deployment in resource-constrained computational platforms. In this context, vector quantization is an appealing framework that expresses multiple parameters using a single code, and has recently achieved state-of-the-art network compression on a range of core vision and natural language processing tasks. Key to the success of vector quantization is deciding which parameter groups should be compressed together. Previous work has relied on heuristics that group the spatial dimension of individual convolutional filters, but a general solution remains unaddressed. This is desirable for pointwise convolutions (which dominate modern architectures), linear layers (which have no notion of spatial dimension), and convolutions (when more than one filter is compressed to the same codeword). In this paper we make the observation that the weights of two adjacent layers can be permuted while expressing the same function. We then establish a connection to rate-distortion theory and search for permutations that result in networks that are easier to compress. Finally, we rely on an annealed quantization algorithm to better compress the network and achieve higher final accuracy. We show results on image classification, object detection, and segmentation, reducing the gap with the uncompressed model by 40 to 70% w.r.t. the current state of the art. We will release code to reproduce all our experiments.

With the growing use of camera devices, the industry has many image datasets that provide more opportunities for collaboration between the machine learning community and industry. However, the sensitive information in the datasets discourages data owners from releasing these datasets. Despite recent research devoted to removing sensitive information from images, they provide neither meaningful privacy-utility trade-off nor provable privacy guarantees. In this study, with the consideration of the perceptual similarity, we propose perceptual indistinguishability (PI) as a formal privacy notion particularly for images. We also propose PI-Net, a privacy-preserving mechanism that achieves image obfuscation with PI guarantee. Our study shows that PI-Net achieves significantly better privacy utility trade-off through public image data.

Visual question answering (VQA) takes an image and a natural-language question as input and returns a natural-language answer. To date, VQA models are primarily assessed by their accuracy on high-level reasoning questions. Nevertheless, Given that perception tasks (e.g., recognizing objects) are the building blocks in the compositional process required by high-level reasoning, there is a demanding need to gain insights into how much of a problem low-level perception is. Inspired by the principles of software metamorphic testing, we introduce MetaVQA, a model-agnostic framework for benchmarking perception capability of VQA models. Given an image i, MetaVQA is able to synthesize a low level perception question q. It then jointly transforms (i, q) to one or a set of sub-questions and sub-images. MetaVQA checks whether the answer to (i, q) satisfies metamorphic relationships (MRs), denoting perception consistency, with the composed answers of transformed questions and images. Violating MRs denotes a failure of answering perception questions. MetaVQA successfully detects over 4.9 million perception failures made by popular VQA models with metamorphic testing. The state-of-the-art VQA models (e.g., the champion of VQA 2020 Challenge) suffer from perception consistency problems. In contrast, the Oscar VQA models, by using anchor points to align questions and images, show generally better consistency in perception tasks. We hope MetaVQA will revitalize interest in enhancing the low-level perceptual abilities of VQA models, a cornerstone of high-level reasoning.