CVPR 2022 Papers

We design deep neural networks (DNNs) and corresponding networks' splittings to distribute DNNs' workload to camera sensors and a centralized aggregator on head-mounted devices to meet system performance targets in inference accuracy and latency under the given hardware resource constraints. To achieve an optimal balance among computation, communication, and performance, a split-aware neural architecture search framework, SplitNets, is introduced to conduct model designing, splitting, and communication reduction simultaneously. We further extend the framework to multi-view systems for learning to fuse inputs from multiple camera sensors with optimal performance and systemic efficiency. We validate SplitNets for single-view system on ImageNet as well as multi-view system on 3D classification, and show that the SplitNets framework achieves state-of-the-art (SOTA) performance and system latency compared with existing approaches.

Variational autoencoders (VAEs) have witnessed great success in performing the compression of image datasets. This success, made possible by the bits-back coding framework, has produced competitive compression performance across many benchmarks. However, despite this, VAE architectures are currently limited by a combination of coding practicalities and compression ratios. That is, not only do state-of-the-art methods, such as normalizing flows, often demonstrate out-performance, but the initial bits required in coding makes single and parallel image compression challenging. To remedy this, we introduce Split Hierarchical Variational Compression (SHVC). SHVC introduces two novelties. Firstly, we propose an efficient autoregressive prior, the autoregressive sub-pixel convolution, that allows a generalisation between per-pixel autoregressions and fully factorised probability models. Secondly, we define our coding framework, the autoregressive initial bits, that flexibly supports parallel coding and avoids -- for the first time -- many of the practicalities commonly associated with bits-back coding. In our experiments, we demonstrate SHVC is able to achieve state-of-the-art compression performance across full-resolution lossless image compression tasks, with up to 100x fewer model parameters than competing VAE approaches.

We present a method for semantically transferring the visual appearance of one natural image to another. Specifically, our goal is to generate an image in which objects in a source structure image are "painted" with the visual appearance of their semantically related objects in a target appearance image. Our method works by training a generator given only a single structure/appearance image pair as input. To integrate semantic information into our framework---a pivotal component in tackling this task---our key idea is to leverage a pre-trained and fixed Vision Transformer (ViT) model which serves as an external semantic prior. Specifically, we derive novel representations of structure and appearance extracted from deep ViT features, untwisting them from the learned self-attention modules. We then establish an objective function that splices the desired structure and appearance representations, interweaving them together in the space of ViT features. Our framework, which we term "Splice", does not involve adversarial training, nor does it require any additional input information such as semantic segmentation or correspondences, and can generate high resolution results, e.g., work in HD. We demonstrate high quality results on a variety of in-the-wild image pairs, under significant variations in the number of objects, their pose and appearance.

Event-based cameras bring a unique capability to tracking, being able to function in challenging real-world conditions as a direct result of their high temporal resolution and high dynamic range. These imagers capture events asynchronously that encode rich temporal and spatial information. However, effectively extracting this information from events remains an open challenge. In this work, we propose a spiking transformer network, STNet, for single object tracking. STNet dynamically extracts and fuses information from both temporal and spatial domains. In particular, the proposed architecture features a transformer module to provide global spatial information and a spiking neural network (SNN) module for extracting temporal cues. The spiking threshold of the SNN module is dynamically adjusted based on the statistical cues of the spatial information, which we find essential in providing robust SNN features. We fuse both feature branches dynamically with a novel cross-domain attention fusion algorithm. Extensive experiments on three event-based datasets, FE240hz, EED and VisEvent validate that the proposed STNet outperforms existing state-of-the-art methods in both tracking accuracy and speed with a significant margin.

Generative Adversarial Network (GAN)-based models have greatly facilitated image synthesis. However, the model performance may be degraded when applied to fine-grained data, due to limited training samples and subtle distinction among categories. Different from generic GANs, we address the issue from a new perspective of discovering and utilizing the underlying structure of real data to explicitly regularize the spatial organization of latent space. To reduce the dependence of generative models on labeled data, we propose a semi-supervised hyper-spherical GAN for class-conditional fine-grained image generation, and our model is referred to as SphericGAN. By projecting random vectors drawn from a prior distribution onto a hyper-sphere, we can model more complex distributions, while at the same time the similarity between the resulting latent vectors depends only on the angle, but not on their magnitudes. On the other hand, we also incorporate a mapping network to map real images onto the hyper-sphere, and match latent vectors with the underlying structure of real data via real-fake cluster alignment. As a result, we obtain a spatially organized latent space, which is useful for capturing class-independent variation factors. The experimental results suggest that our SphericGAN achieves state-of-the-art performance in synthesizing high-fidelity images with precise class semantics.

The 360deg imaging has recently gained much attention; however, its angular resolution is relatively lower than that of a narrow field-of-view (FOV) perspective image as it is captured using a fisheye lens with the same sensor size. Therefore, it is beneficial to super-resolve a 360deg image. Several attempts have been made, but mostly considered equirectangular projection (ERP) as one of the ways for 360deg image representation despite the latitude-dependent distortions. In that case, as the output high-resolution (HR) image is always in the same ERP format as the lowresolution (LR) input, additional information loss may occur when transforming the HR image to other projection types. In this paper, we propose SphereSR, a novel framework to generate a continuous spherical image representation from an LR 360deg image, with the goal of predicting the RGB values at given spherical coordinates for superresolution with an arbitrary 360deg image projection. Specifically, first we propose a feature extraction module that represents the spherical data based on an icosahedron and that efficiently extracts features on the spherical surface. We then propose a spherical local implicit image function (SLIIF) to predict RGB values at the spherical coordinates. As such, SphereSR flexibly reconstructs an HR image given an arbitrary projection type. Experiments on various benchmark datasets show that the proposed method significantly surpasses existing methods in terms of performance.

With the appearance of super high-resolution (e.g., gigapixel-level) images, performing efficient object detection on such images becomes an important issue. Most existing works for efficient object detection on high-resolution images focus on generating local patches where objects may exist, and then every patch is detected independently. However, when the image resolution reaches gigapixel-level, they will suffer from a huge time cost for detecting numerous patches. Different from them, we devise a novel patch arrangement framework for fast object detection on gigapixel-level images. Under this framework, a Patch Arrangement Network (PAN) is proposed to accelerate the detection by determining which patches could be packed together into a compact canvas. Specifically, PAN consists of (1) a Patch Filter Module (PFM) (2) a Patch Packing Module (PPM). PFM filters patch candidates by learning to select patches between two granularities. Subsequently, from the remaining patches, PPM determines how to pack these patches together into a smaller number of canvases. Meanwhile, it generates an ideal layout of patches on canvas. These canvases are fed to the detector to get final results. Experiments show that our method could improve the inference speed on gigapixel-level images by 5 times while maintaining great performance.

Advances in speech driven animation techniques allow the creation of convincing animations for virtual characters solely from audio data. Many existing approaches focus on facial and lip motion and they often do not provide realistic animation of the inner mouth. This paper addresses the problem of speech-driven inner mouth animation. Obtaining performance capture data of the tongue and jaw from video alone is difficult because the inner mouth is only partially observable during speech. In this work, we introduce a large-scale speech and mocap dataset that focuses on capturing tongue, jaw, and lip motion. This dataset enables research using data-driven techniques to generate realistic inner mouth animation from speech. We then propose a deep-learning based method for accurate and generalizable speech to tongue and jaw animation and evaluate several encoder-decoder network architectures and audio feature encoders. We find that recent self-supervised deep learning based audio feature encoders are robust, generalize well to unseen speakers and content, and work best for our task. To demonstrate the practical application of our approach, we show animations on high-quality parametric 3D face models driven by the landmarks generated from our speech-to-tongue animation method.

Though unsupervised domain adaptation (UDA) has achieved very impressive progress recently, it remains a great challenge due to missing target annotations and the rich discrepancy between source and target distributions. We propose Spectral UDA (SUDA), an effective and efficient UDA technique that works in the spectral space and can generalize across different visual recognition tasks. SUDA addresses the UDA challenges from two perspectives. First, it introduces a spectrum transformer (ST) that mitigates inter-domain discrepancies by enhancing domain-invariant spectra while suppressing domain-variant spectra of source and target samples simultaneously. Second, it introduces multi-view spectral learning that learns useful unsupervised representations by maximizing mutual information among multiple ST-generated spectral views of each target sample. Extensive experiments show that SUDA achieves superior accuracy consistently across different visual tasks in image classification, semantic segmentation, and object detection. Additionally, SUDA also works with the transformer-based network and achieves state-of-the-art performance on object detection.

We propose a novel few-shot action recognition framework, STRM, which enhances class-specific feature discriminability while simultaneously learning higher-order temporal representations. The focus of our approach is a novel spatio-temporal enrichment module that aggregates spatial and temporal contexts with dedicated local patch-level and global frame-level feature enrichment sub-modules. Local patch-level enrichment captures the appearance-based characteristics of actions. On the other hand, global frame-level enrichment explicitly encodes the broad temporal context, thereby capturing the relevant object features over time. The resulting spatio-temporally enriched representations are then utilized to learn the relational matching between query and support action sub-sequences. We further introduce a query-class similarity classifier on the patch-level enriched features to enhance class-specific feature discriminability by reinforcing the feature learning at different stages in the proposed framework. Experiments are performed on four few-shot action recognition benchmarks: Kinetics, SSv2, HMDB51 and UCF101. Our extensive ablation study reveals the benefits of the proposed contributions. Furthermore, our approach sets a new state-of-the-art on all four benchmarks. On the challenging SSv2 benchmark, our approach achieves an absolute gain of 3.5% in classification accuracy, as compared to the best existing method in the literature. Our code and models are available at https://github.com/Anirudh257/strm.

Predicting future motion based on historical motion sequence is a fundamental problem in computer vision, and it has wide applications in autonomous driving and robotics. Some recent works have shown that Graph Convolutional Networks(GCN) are instrumental in modeling the relationship between different joints. However, considering the variants and diverse action types in human motion data, the cross-dependency of the spatio-temporal relationships will be difficult to depict due to the decoupled modeling strategy, which may also exacerbate the problem of insufficient generalization. Therefore, we propose the Spatio-Temporal Gating-Adjacency GCN(GAGCN) to learn the complex spatio-temporal dependencies over diverse action types. Specifically, we adopt gating networks to enhance the generalization of GCN via the trainable adaptive adjacency matrix obtained by blending the candidate spatio-temporal adjacency matrices. Moreover, GAGCN addresses the cross-dependency of space and time by balancing the weights of spatio-temporal modeling and fusing the decoupled spatio-temporal features. Extensive experiments on Human 3.6M, AMASS, and 3DPW demonstrate that GAGCN achieves state-of-the-art performance in both short-term and long-term predictions.

Existing GAN inversion and editing methods work well for aligned objects with a clean background, such as portraits and animal faces, but often struggle for more difficult categories with complex scene layouts and object occlusions, such as cars, animals, and outdoor images. We propose a new method to invert and edit such complex images in the latent space of GANs, such as StyleGAN2. Our key idea is to explore inversion with a collection of layers, spatially adapting the inversion process to the difficulty of the image. We learn to predict the "invertibility" of different image segments and project each segment into a latent layer. Easier regions can be inverted into an earlier layer in the generator's latent space, while more challenging regions can be inverted into a later feature space. Experiments show that our method obtains better inversion results compared to the recent approaches on complex categories, while maintaining downstream editability. Please refer to our project page at gauravparmar.com/sam_ inversion.

Spatial-Temporal Video Super-Resolution (ST-VSR) aims to generate super-resolved videos with higher resolution (HR) and higher frame rate (HFR). Quite intuitively, pioneering two-stage based methods complete ST-VSR directly combining two sub-tasks: Spatial Video Super-Resolution (S-VSR) and Temporal Video Super-Resolution (T-VSR) but ignore the reciprocal relations among them. Specifically, 1) T-VSR to S-VSR: temporal correlations help accurate spatial detail representation with more clues; 2) S-VSR to T-VSR: abundant spatial information contributes to the refinement of temporal prediction. To this end, we propose a one-stage based Cycle-projected Mutual learning network (CycMu-Net) for ST-VSR, which makes full use of spatial-temporal correlations via the mutual learning between S-VSR and T-VSR. Specifically, we propose to exploit the mutual information among them via iterative up-and-down projections, where the spatial and temporal features are fully fused and distilled, helping the high-quality video reconstruction. Besides extensive experiments on benchmark datasets, we also compare our proposed CycMu-Net with S-VSR and T-VSR tasks, demonstrating that our method significantly outperforms state-of-the-art methods. Codes are publicly available at: https://github.com/hhhhhumengshun/CycMuNet.

We propose an approach to estimate arm and hand dynamics from monocular video by utilizing the relationship between arm and hand. Although monocular full human motion capture technologies have made great progress in recent years, recovering accurate and plausible arm twists and hand gestures from in-the-wild videos still remains a challenge. To solve this problem, our solution is proposed based on the fact that arm poses and hand gestures are highly correlated in most real situations. To make best use of arm-hand correlation as well as inter-frame information, we carefully design a Spatial-Temporal Parallel Arm-Hand Motion Transformer (PAHMT) to predict the arm and hand dynamics simultaneously. We also introduce new losses to encourage the estimations to be smooth and accurate. Besides, we collect a motion capture dataset including 200K frames of hand gestures and use this data to train our model. By integrating a 2D hand pose estimation model and a 3D human pose estimation model, the proposed method can produce plausible arm and hand dynamics from monocular video. Extensive evaluations demonstrate that the proposed method has advantages over previous state-of-the-art approaches and shows robustness under various challenging scenarios.

We solve object localisation in partial scenes, a new problem of estimating the unknown position of an object (e.g. where is the bag?) given a partial 3D scan of a scene. The proposed solution is based on a novel scene graph model, the Spatial Commonsense Graph (SCG), where objects are the nodes and edges define pairwise distances between them, enriched by concept nodes and relationships from a commonsense knowledge base. This allows SCG to better generalise its spatial inference to unknown 3D scenes. The SCG is used to estimate the unknown position of the target object in two steps: first, we feed the SCG into a novel Proximity Prediction Network, a graph neural network that uses attention to perform distance prediction between the node representing the target object and the nodes representing the observed objects in the SCG; second, we propose a Localisation Module based on circular intersection to estimate the object position using all the predicted pairwise distances in order to be independent of any reference system. We create a new dataset of partially reconstructed scenes to benchmark our method and baselines for object localisation in partial scenes, where our proposed method achieves the best localisation performance. Code and Dataset are available here: https://github.com/IIT-PAVIS/SpatialCommonsenseGraph

In this paper, we propose a solution to the task of generating dynamic 3D facial expressions from a neutral 3D face and an expression label. This involves solving two sub-problems: (i) modeling the temporal dynamics of expressions, and (ii) deforming the neutral mesh to obtain the expressive counterpart. We represent the temporal evolution of expressions using the motion of a sparse set of 3D landmarks that we learn to generate by training a manifold-valued GAN (Motion3DGAN). To better encode the expression-induced deformation and disentangle it from the identity information, the generated motion is represented as per-frame displacement from a neutral configuration. To generate the expressive meshes, we train a Sparse2Dense mesh Decoder (S2D-Dec) that maps the landmark displacements to a dense, per-vertex displacement. This allows us to learn how the motion of a sparse set of landmarks influences the deformation of the overall face surface, independently from the identity. Experimental results on the CoMA and D3DFACS datasets show that our solution brings significant improvements with respect to previous solutions in terms of both dynamic expression generation and mesh reconstruction, while retaining good generalization to unseen data. The code and the pretrained model will be made publicly available.

Geospatial semantic segmentation on remote sensing images suffers from large intra-class variance in both foreground and background classes. First, foreground objects are tiny in the remote sensing images and are represented by only a few pixels, which leads to large foreground intra-class variance and undermines the discrimination between foreground classes (issue firstly considered in this work). Second, background class contains complex context, which results in false alarms due to large background intra-class variance. To alleviate these two issues, we construct a sparse and complete latent structure via prototypes. In particular, to enhance the sparsity of the latent space, we design a prototypical contrastive learning to have prototypes of the same category clustering together and prototypes of different categories to be far away from each other. Also, we strengthen the completeness of the latent space by modeling all foreground categories and hardest (nearest) background objects. We further design a patch shuffle augmentation for remote sensing images with complicated contexts. Our augmentation encourages the semantic information of an object to be correlated only to the limited context within the patch that is specific to its category, which further reduces large intra-class variance. We conduct extensive evaluations on a large scale remote sensing dataset, showing our approach significantly outperforms state-of-the-art methods by a large margin.

Most state-of-the-art instance segmentation methods have to be trained on densely annotated images. While difficult in general, this requirement is especially daunting for biomedical images, where domain expertise is often required for annotation and no large public data collections are available for pre-training. We propose to address the dense annotation bottleneck by introducing a proposal-free segmentation approach based on non-spatial embeddings, which exploits the structure of the learned embedding space to extract individual instances in a differentiable way. The segmentation loss can then be applied directly to instances and the overall pipeline can be trained in a fully- or weakly supervised manner. We consider the challenging case of positive-unlabeled supervision, where a novel self-supervised consistency loss is introduced for the unlabeled parts of the training data. We evaluate the proposed method on 2D and 3D segmentation problems in different microscopy modalities as well as on the Cityscapes and CVPPP instance segmentation benchmarks, achieving state-of-the-art results on the latter.

CRF is a classical computer vision model which is also useful for deep learning. There are two common CRF types: sparse and dense. Sparse CRF connects only the nearby pixels, while dense CRF has global connectivity. Therefore dense CRF is a more general model, but it is much harder to optimize compared to sparse CRF. In fact, only a certain form of dense CRF is optimized in practice, and even then approximately. We propose a new sparse non-local CRF: it has a sparse number of connections, but it has both local and non-local ones. Like sparse CRF, the total number of connections is small, and our model is easy to optimize exactly. Like dense CRF, our model is more general than sparse CRF due to non-local connections. We show that our sparse non-local CRF can model properties similar to that of the popular Gaussian edge dense CRF. Besides efficiency, another advantage is that our edge weights are less restricted compared to Gaussian edge dense CRF. We design models that take advantage of this flexibility. We also discuss connection of our model to other CRF models. Finally, to prove the usefulness of our model, we evaluate it on the classical application of segmentation from a bounding box and for deep learning based salient object segmentation. We improve state of the art for both applications.