CVPR 2023 Papers

Different video understanding tasks are typically treated in isolation, and even with distinct types of curated data (e.g., classifying sports in one dataset, tracking animals in another). However, in wearable cameras, the immersive egocentric perspective of a person engaging with the world around them presents an interconnected web of video understanding tasks---hand-object manipulations, navigation in the space, or human-human interactions---that unfold continuously, driven by the person's goals. We argue that this calls for a much more unified approach. We propose EgoTask Translation (EgoT2), which takes a collection of models optimized on separate tasks and learns to translate their outputs for improved performance on any or all of them at once. Unlike traditional transfer or multi-task learning, EgoT2's "flipped design" entails separate task-specific backbones and a task translator shared across all tasks, which captures synergies between even heterogeneous tasks and mitigates task competition. Demonstrating our model on a wide array of video tasks from Ego4D, we show its advantages over existing transfer paradigms and achieve top-ranked results on four of the Ego4D 2022 benchmark challenges.

In a noisy conversation environment such as a dinner party, people often exhibit selective auditory attention, or the ability to focus on a particular speaker while tuning out others. Recognizing who somebody is listening to in a conversation is essential for developing technologies that can understand social behavior and devices that can augment human hearing by amplifying particular sound sources. The computer vision and audio research communities have made great strides towards recognizing sound sources and speakers in scenes. In this work, we take a step further by focusing on the problem of localizing auditory attention targets in egocentric video, or detecting who in a camera wearer's field of view they are listening to. To tackle the new and challenging Selective Auditory Attention Localization problem, we propose an end-to-end deep learning approach that uses egocentric video and multichannel audio to predict the heatmap of the camera wearer's auditory attention. Our approach leverages spatiotemporal audiovisual features and holistic reasoning about the scene to make predictions, and outperforms a set of baselines on a challenging multi-speaker conversation dataset. Project page: https://fkryan.github.io/saal

Humans naturally perceive surrounding scenes by unifying sound and sight in a first-person view. Likewise, machines are advanced to approach human intelligence by learning with multisensory inputs from an egocentric perspective. In this paper, we explore the challenging egocentric audio-visual object localization task and observe that 1) egomotion commonly exists in first-person recordings, even within a short duration; 2) The out-of-view sound components can be created while wearers shift their attention. To address the first problem, we propose a geometry-aware temporal aggregation module to handle the egomotion explicitly. The effect of egomotion is mitigated by estimating the temporal geometry transformation and exploiting it to update visual representations. Moreover, we propose a cascaded feature enhancement module to tackle the second issue. It improves cross-modal localization robustness by disentangling visually-indicated audio representation. During training, we take advantage of the naturally available audio-visual temporal synchronization as the "free" self-supervision to avoid costly labeling. We also annotate and create the Epic Sounding Object dataset for evaluation purposes. Extensive experiments show that our method achieves state-of-the-art localization performance in egocentric videos and can be generalized to diverse audio-visual scenes.

Estimating 3D human motion from an egocentric video sequence plays a critical role in human behavior understanding and has various applications in VR/AR. However, naively learning a mapping between egocentric videos and human motions is challenging, because the user's body is often unobserved by the front-facing camera placed on the head of the user. In addition, collecting large-scale, high-quality datasets with paired egocentric videos and 3D human motions requires accurate motion capture devices, which often limit the variety of scenes in the videos to lab-like environments. To eliminate the need for paired egocentric video and human motions, we propose a new method, Ego-Body Pose Estimation via Ego-Head Pose Estimation (EgoEgo), which decomposes the problem into two stages, connected by the head motion as an intermediate representation. EgoEgo first integrates SLAM and a learning approach to estimate accurate head motion. Subsequently, leveraging the estimated head pose as input, EgoEgo utilizes conditional diffusion to generate multiple plausible full-body motions. This disentanglement of head and body pose eliminates the need for training datasets with paired egocentric videos and 3D human motion, enabling us to leverage large-scale egocentric video datasets and motion capture datasets separately. Moreover, for systematic benchmarking, we develop a synthetic dataset, AMASS-Replica-Ego-Syn (ARES), with paired egocentric videos and human motion. On both ARES and real data, our EgoEgo model performs significantly better than the current state-of-the-art methods.

Vision transformers have shown great success due to their high model capabilities. However, their remarkable performance is accompanied by heavy computation costs, which makes them unsuitable for real-time applications. In this paper, we propose a family of high-speed vision transformers named EfficientViT. We find that the speed of existing transformer models is commonly bounded by memory inefficient operations, especially the tensor reshaping and element-wise functions in MHSA. Therefore, we design a new building block with a sandwich layout, i.e., using a single memory-bound MHSA between efficient FFN layers, which improves memory efficiency while enhancing channel communication. Moreover, we discover that the attention maps share high similarities across heads, leading to computational redundancy. To address this, we present a cascaded group attention module feeding attention heads with different splits of the full feature, which not only saves computation cost but also improves attention diversity. Comprehensive experiments demonstrate EfficientViT outperforms existing efficient models, striking a good trade-off between speed and accuracy. For instance, our EfficientViT-M5 surpasses MobileNetV3-Large by 1.9% in accuracy, while getting 40.4% and 45.2% higher throughput on Nvidia V100 GPU and Intel Xeon CPU, respectively. Compared to the recent efficient model MobileViT-XXS, EfficientViT-M2 achieves 1.8% superior accuracy, while running 5.8x/3.7x faster on the GPU/CPU, and 7.4x faster when converted to ONNX format. Code and models will be available soon.

Video snapshot compressive imaging (SCI) uses a two-dimensional detector to capture consecutive video frames during a single exposure time. Following this, an efficient reconstruction algorithm needs to be designed to reconstruct the desired video frames. Although recent deep learning-based state-of-the-art (SOTA) reconstruction algorithms have achieved good results in most tasks, they still face the following challenges due to excessive model complexity and GPU memory limitations: 1) these models need high computational cost, and 2) they are usually unable to reconstruct large-scale video frames at high compression ratios. To address these issues, we develop an efficient network for video SCI by using dense connections and space-time factorization mechanism within a single residual block, dubbed EfficientSCI. The EfficientSCI network can well establish spatial-temporal correlation by using convolution in the spatial domain and Transformer in the temporal domain, respectively. We are the first time to show that an UHD color video with high compression ratio can be reconstructed from a snapshot 2D measurement using a single end-to-end deep learning model with PSNR above 32 dB. Extensive results on both simulation and real data show that our method significantly outperforms all previous SOTA algorithms with better real-time performance.

The aim of this paper is to propose a mechanism to efficiently and explicitly model image hierarchies in the global, regional, and local range for image restoration. To achieve that, we start by analyzing two important properties of natural images including cross-scale similarity and anisotropic image features. Inspired by that, we propose the anchored stripe self-attention which achieves a good balance between the space and time complexity of self-attention and the modelling capacity beyond the regional range. Then we propose a new network architecture dubbed GRL to explicitly model image hierarchies in the Global, Regional, and Local range via anchored stripe self-attention, window self-attention, and channel attention enhanced convolution. Finally, the proposed network is applied to 7 image restoration types, covering both real and synthetic settings. The proposed method sets the new state-of-the-art for several of those. Code will be available at https://github.com/ofsoundof/GRL-Image-Restoration.git.

While current multi-frame restoration methods combine information from multiple input images using 2D alignment techniques, recent advances in novel view synthesis are paving the way for a new paradigm relying on volumetric scene representations. In this work, we introduce the first 3D-based multi-frame denoising method that significantly outperforms its 2D-based counterparts with lower computational requirements. Our method extends the multiplane image (MPI) framework for novel view synthesis by introducing a learnable encoder-renderer pair manipulating multiplane representations in feature space. The encoder fuses information across views and operates in a depth-wise manner while the renderer fuses information across depths and operates in a view-wise manner. The two modules are trained end-to-end and learn to separate depths in an unsupervised way, giving rise to Multiplane Feature (MPF) representations. Experiments on the Spaces and Real Forward-Facing datasets as well as on raw burst data validate our approach for view synthesis, multi-frame denoising, and view synthesis under noisy conditions.

The deployment of perception systems based on neural networks in safety critical applications requires assurance on their robustness. Deterministic guarantees on network robustness require formal verification. Standard approaches for verifying robustness analyse invariance to analytically defined transformations, but not the diverse and ubiquitous changes involving object pose, scene viewpoint, occlusions, etc. To this end, we present an efficient approach for verifying specifications definable using Latent Variable Models that capture such diverse changes. The approach involves adding an invertible encoding head to the network to be verified, enabling the verification of latent space sets with minimal reconstruction overhead. We report verification experiments for three classes of proposed latent space specifications, each capturing different types of realistic input variations. Differently from previous work in this area, the proposed approach is relatively independent of input dimensionality and scales to a broad class of deep networks and real-world datasets by mitigating the inefficiency and decoder expressivity dependence in the present state-of-the-art.

Video semantic segmentation (VSS) is a computationally expensive task due to the per-frame prediction for videos of high frame rates. In recent work, compact models or adaptive network strategies have been proposed for efficient VSS. However, they did not consider a crucial factor that affects the computational cost from the input side: the input resolution. In this paper, we propose an altering resolution framework called AR-Seg for compressed videos to achieve efficient VSS. AR-Seg aims to reduce the computational cost by using low resolution for non-keyframes. To prevent the performance degradation caused by downsampling, we design a Cross Resolution Feature Fusion (CReFF) module, and supervise it with a novel Feature Similarity Training (FST) strategy. Specifically, CReFF first makes use of motion vectors stored in a compressed video to warp features from high-resolution keyframes to low-resolution non-keyframes for better spatial alignment, and then selectively aggregates the warped features with local attention mechanism. Furthermore, the proposed FST supervises the aggregated features with high-resolution features through an explicit similarity loss and an implicit constraint from the shared decoding layer. Extensive experiments on CamVid and Cityscapes show that AR-Seg achieves state-of-the-art performance and is compatible with different segmentation backbones. On CamVid, AR-Seg saves 67% computational cost (measured in GFLOPs) with the PSPNet18 backbone while maintaining high segmentation accuracy. Code: https://github.com/THU-LYJ-Lab/AR-Seg.

Planes are generally used in 3D reconstruction for depth sensors, such as RGB-D cameras and LiDARs. This paper focuses on the problem of estimating the optimal planes and sensor poses to minimize the point-to-plane distance. The resulting least-squares problem is referred to as plane adjustment (PA) in the literature, which is the counterpart of bundle adjustment (BA) in visual reconstruction. Iterative methods are adopted to solve these least-squares problems. Typically, Newton's method is rarely used for a large-scale least-squares problem, due to the high computational complexity of the Hessian matrix. Instead, methods using an approximation of the Hessian matrix, such as the Levenberg-Marquardt (LM) method, are generally adopted. This paper adopts the Newton's method to efficiently solve the PA problem. Specifically, given poses, the optimal plane have a close-form solution. Thus we can eliminate planes from the cost function, which significantly reduces the number of variables. Furthermore, as the optimal planes are functions of poses, this method actually ensures that the optimal planes for the current estimated poses can be obtained at each iteration, which benefits the convergence. The difficulty lies in how to efficiently compute the Hessian matrix and the gradient of the resulting cost. This paper provides an efficient solution. Empirical evaluation shows that our algorithm outperforms the state-of-the-art algorithms.

Any-scale image synthesis offers an efficient and scalable solution to synthesize photo-realistic images at any scale, even going beyond 2K resolution. However, existing GAN-based solutions depend excessively on convolutions and a hierarchical architecture, which introduce inconsistency and the "texture sticking" issue when scaling the output resolution. From another perspective, INR-based generators are scale-equivariant by design, but their huge memory footprint and slow inference hinder these networks from being adopted in large-scale or real-time systems. In this work, we propose Column-Row Entangled Pixel Synthesisthes (CREPS), a new generative model that is both efficient and scale-equivariant without using any spatial convolutions or coarse-to-fine design. To save memory footprint and make the system scalable, we employ a novel bi-line representation that decomposes layer-wise feature maps into separate "thick" column and row encodings. Experiments on standard datasets, including FFHQ, LSUN-Church, and MetFaces, confirm CREPS' ability to synthesize scale-consistent and alias-free images up to 4K resolution with proper training and inference speed.

Robust principal component analysis (RPCA) is widely studied in computer vision. Recently an adaptive rank estimate based RPCA has achieved top performance in low-level vision tasks without the prior rank, but both the rank estimate and RPCA optimization algorithm involve singular value decomposition, which requires extremely huge computational resource for large-scale matrices. To address these issues, an efficient RPCA (eRPCA) algorithm is proposed based on block Krylov iteration and CUR decomposition in this paper. Specifically, the Krylov iteration method is employed to approximate the eigenvalue decomposition in the rank estimation, which requires O(ndrq + n(rq)^2) for an (nxd) input matrix, in which q is a parameter with a small value, r is the target rank. Based on the estimated rank, CUR decomposition is adopted to replace SVD in updating low-rank matrix component, whose complexity reduces from O(rnd) to O(r^2n) per iteration. Experimental results verify the efficiency and effectiveness of the proposed eRPCA over the state-of-the-art methods in various low-level vision applications.

Most current RGB-T trackers adopt a two-stream structure to extract unimodal RGB and thermal features and complex fusion strategies to achieve multi-modal feature fusion, which require a huge number of parameters, thus hindering their real-life applications. On the other hand, a compact RGB-T tracker may be computationally efficient but encounter non-negligible performance degradation, due to the weakening of feature representation ability. To remedy this situation, a cross-modality distillation framework is presented to bridge the performance gap between a compact tracker and a powerful tracker. Specifically, a specific-common feature distillation module is proposed to transform the modality-common information as well as the modality-specific information from a deeper two-stream network to a shallower single-stream network. In addition, a multi-path selection distillation module is proposed to instruct a simple fusion module to learn more accurate multi-modal information from a well-designed fusion mechanism by using multiple paths. We validate the effectiveness of our method with extensive experiments on three RGB-T benchmarks, which achieves state-of-the-art performance but consumes much less computational resources.

Despite its importance for federated learning, continuous learning and many other applications, on-device training remains an open problem for EdgeAI. The problem stems from the large number of operations (e.g., floating point multiplications and additions) and memory consumption required during training by the back-propagation algorithm. Consequently, in this paper, we propose a new gradient filtering approach which enables on-device CNN model training. More precisely, our approach creates a special structure with fewer unique elements in the gradient map, thus significantly reducing the computational complexity and memory consumption of back propagation during training. Extensive experiments on image classification and semantic segmentation with multiple CNN models (e.g., MobileNet, DeepLabV3, UPerNet) and devices (e.g., Raspberry Pi and Jetson Nano) demonstrate the effectiveness and wide applicability of our approach. For example, compared to SOTA, we achieve up to 19x speedup and 77.1% memory savings on ImageNet classification with only 0.1% accuracy loss. Finally, our method is easy to implement and deploy; over 20x speedup and 90% energy savings have been observed compared to highly optimized baselines in MKLDNN and CUDNN on NVIDIA Jetson Nano. Consequently, our approach opens up a new direction of research with a huge potential for on-device training.

Large-scale pre-training has brought unimodal fields such as computer vision and natural language processing to a new era. Following this trend, the size of multimodal learning models constantly increases, leading to an urgent need to reduce the massive computational cost of fine-tuning these models for downstream tasks. In this paper, we propose an efficient and flexible multimodal fusion method, namely PMF, tailored for fusing unimodally pretrained transformers. Specifically, we first present a modular multimodal fusion framework that exhibits high flexibility and facilitates mutual interactions among different modalities. In addition, we disentangle vanilla prompts into three types in order to learn different optimizing objectives for multimodal learning. It is also worth noting that we propose to add prompt vectors only on the deep layers of the unimodal transformers, thus significantly reducing the training memory usage. Experiment results show that our proposed method achieves comparable performance to several other multimodal finetuning methods with less than 3% trainable parameters and up to 66% saving of training memory usage.

The ability to distinguish between different movie scenes is critical for understanding the storyline of a movie. However, accurately detecting movie scenes is often challenging as it requires the ability to reason over very long movie segments. This is in contrast to most existing video recognition models, which are typically designed for short-range video analysis. This work proposes a State-Space Transformer model that can efficiently capture dependencies in long movie videos for accurate movie scene detection. Our model, dubbed TranS4mer, is built using a novel S4A building block, which combines the strengths of structured state-space sequence (S4) and self-attention (A) layers. Given a sequence of frames divided into movie shots (uninterrupted periods where the camera position does not change), the S4A block first applies self-attention to capture short-range intra-shot dependencies. Afterward, the state-space operation in the S4A block is used to aggregate long-range inter-shot cues. The final TranS4mer model, which can be trained end-to-end, is obtained by stacking the S4A blocks one after the other multiple times. Our proposed TranS4mer outperforms all prior methods in three movie scene detection datasets, including MovieNet, BBC, and OVSD, while also being 2x faster and requiring 3x less GPU memory than standard Transformer models. We will release our code and models.

The goal of click-based interactive image segmentation is to extract target masks with the input of positive/negative clicks. Every time a new click is placed, existing methods run the whole segmentation network to obtain a corrected mask, which is inefficient since several clicks may be needed to reach satisfactory accuracy. To this end, we propose an efficient method to correct the mask with a lightweight mask correction network. The whole network remains a low computational cost from the second click, even if we have a large backbone. However, a simple correction network with limited capacity is not likely to achieve comparable performance with a classic segmentation network. Thus, we propose a click-guided self-attention module and a click-guided correlation module to effectively exploits the click information to boost performance. First, several templates are selected based on the semantic similarity with click features. Then the self-attention module propagates the template information to other pixels, while the correlation module directly uses the templates to obtain target outlines. With the efficient architecture and two click-guided modules, our method shows preferable performance and efficiency compared to existing methods. The code will be released at https://github.com/feiaxyt/EMC-Click.

Localization in a pre-built map is a basic technique for robot autonomous navigation. Existing mapping and localization methods commonly work well in small-scale environments. As a map grows larger, however, more memory is required and localization becomes inefficient. To solve these problems, map sparsification becomes a practical necessity to acquire a subset of the original map for localization. Previous map sparsification methods add a quadratic term in mixed-integer programming to enforce a uniform distribution of selected landmarks, which requires high memory capacity and heavy computation. In this paper, we formulate map sparsification in an efficient linear form and select uniformly distributed landmarks based on 2D discretized grids. Furthermore, to reduce the influence of different spatial distributions between the mapping and query sequences, which is not considered in previous methods, we also introduce a space constraint term based on 3D discretized grids. The exhaustive experiments in different datasets demonstrate the superiority of the proposed methods in both efficiency and localization performance. The relevant codes will be released at https://github.com/fishmarch/SLAM_Map_Compression.