CVPR 2022 论文列表

Catastrophic forgetting (CF) frequently occurs when learning with non-stationary data distribution. The CF issue remains nearly unexplored and is more challenging when meta-learning on a sequence of domains (datasets), called sequential domain meta-learning (SDML). In this work, we propose a simple yet effective learning to learn approach, i.e., meta optimizer, to mitigate the CF problem in SDML. We first apply the proposed meta optimizer to the simplified setting of SDML, domain-aware meta-learning, where the domain labels and boundaries are known during the learning process. We propose dynamically freezing the network and incorporating it with the proposed meta optimizer by considering the domain nature during meta training. In addition, we extend the meta optimizer to the more general setting of SDML, domain-agnostic meta-learning, where domain labels and boundaries are unknown during the learning process. We propose a domain shift detection technique to capture latent domain change and equip the meta optimizer with it to work in this setting. The proposed meta optimizer is versatile and can be easily integrated with several existing meta-learning algorithms. Finally, we construct a challenging and large-scale benchmark consisting of 10 heterogeneous domains with a super long task sequence consisting of 100K tasks. We perform extensive experiments on the proposed benchmark for both settings and demonstrate the effectiveness of our proposed method, outperforming current strong baselines by a large margin.

Modern camera localization methods that use image retrieval, feature matching, and 3D structure-based pose estimation require long-term storage of numerous scene images or a vast amount of image features. This can make them unsuitable for resource constrained VR/AR devices and also raises serious privacy concerns. We present a new learned camera localization technique that eliminates the need to store features or a detailed 3D point cloud. Our key idea is to implicitly encode the appearance of a sparse yet salient set of 3D scene points into a convolutional neural network (CNN) that can detect these scene points in query images whenever they are visible. We refer to these points as scene landmarks. We also show that a CNN can be trained to regress bearing vectors for such landmarks even when they are not within the camera's field-of-view. We demonstrate that the predicted landmarks yield accurate pose estimates and that our method outperforms DSAC*, the state-of-the-art in learned localization. Furthermore, extending HLoc (an accurate method) by combining its correspondences with our predictions, boosts its accuracy even further.

In this paper, we focus on the Audio-Visual Question Answering (AVQA) task, which aims to answer questions regarding different visual objects, sounds, and their associations in videos. The problem requires comprehensive multimodal understanding and spatio-temporal reasoning over audio-visual scenes. To benchmark this task and facilitate our study, we introduce a large-scale MUSIC-AVQA dataset, which contains more than 45K question-answer pairs covering 33 different question templates spanning over different modalities and question types. We develop several baselines and introduce a spatio-temporal grounded audio-visual network for the AVQA problem. Our results demonstrate that AVQA benefits from multisensory perception and our model outperforms recent A-, V-, and AVQA approaches. We believe that our built dataset has the potential to serve as testbed for evaluating and promoting progress in audio-visual scene understanding and spatio-temporal reasoning. Code and dataset: http://gewu-lab.github.io/MUSIC-AVQA/

In this paper, we are interested in modeling a how-to instructional procedure, such as a cooking recipe, with a meaningful and rich high-level representation. Specifically, we propose to represent cooking recipes and food images as cooking programs. Programs provide a structured representation of the task, capturing cooking semantics and sequential relationships of actions in the form of a graph. This allows them to be easily manipulated by users and executed by agents. To this end, we build a model that is trained to learn a joint embedding between recipes and food images via self-supervision and jointly generate a program from this embedding as a sequence. To validate our idea, we crowdsource programs for cooking recipes and show that: (a) projecting the image-recipe embeddings into programs leads to better cross-modal retrieval results; (b) generating programs from images leads to better recognition results compared to predicting raw cooking instructions; and (c) we can generate food images by manipulating programs via optimizing the latent code of a GAN. Code, data, and models are available online.

Part segmentations provide a rich and detailed part-level description of objects. However, their annotation requires an enormous amount of work, which makes it difficult to apply standard deep learning methods. In this paper, we propose the idea of learning part segmentation through unsupervised domain adaptation (UDA) from synthetic data. We first introduce UDA-Part, a comprehensive part segmentation dataset for vehicles that can serve as an adequate benchmark for UDA (https://qliu24.github.io/udapart/). In UDA-Part, we label parts on 3D CAD models which enables us to generate a large set of annotated synthetic images. We also annotate parts on a number of real images to provide a real test set. Secondly, to advance the adaptation of part models trained from the synthetic data to the real images, we introduce a new UDA algorithm that leverages the object's spatial structure to guide the adaptation process. Our experimental results on two real test datasets confirm the superiority of our approach over existing works, and demonstrate the promise of learning part segmentation for general objects from synthetic data. We believe our dataset provides a rich testbed to study UDA for part segmentation and will help to significantly push forward research in this area.

Despite the recent advances in multi-task learning of dense prediction problems, most methods rely on expensive labelled datasets. In this paper, we present a label efficient approach and look at jointly learning of multiple dense prediction tasks on partially annotated data (i.e. not all the task labels are available for each image), which we call multi-task partially-supervised learning. We propose a multi-task training procedure that successfully leverages task relations to supervise its multi-task learning when data is partially annotated. In particular, we learn to map each task pair to a joint pairwise task-space which enables sharing information between them in a computationally efficient way through another network conditioned on task pairs, and avoids learning trivial cross-task relations by retaining high-level information about the input image. We rigorously demonstrate that our proposed method effectively exploits the images with unlabelled tasks and outperforms existing semi-supervised learning approaches and related methods on three standard benchmarks.

Recent works on 3D semantic segmentation propose to exploit the synergy between images and point clouds by processing each modality with a dedicated network and projecting learned 2D features onto 3D points. Merging large-scale point clouds and images raises several challenges, such as constructing a mapping between points and pixels, and aggregating features between multiple views. Current methods require mesh reconstruction or specialized sensors to recover occlusions, and use heuristics to select and aggregate available images. In contrast, we propose an end-to-end trainable multi-view aggregation model leveraging the viewing conditions of 3D points to merge features from images taken at arbitrary positions. Our method can combine standard 2D and 3D networks and outperforms both 3D models operating on colorized point clouds and hybrid 2D/3D networks without requiring colorization, meshing, or true depth maps. We set a new state-of-the-art for large-scale indoor/outdoor semantic segmentation on S3DIS (74.7 mIoU 6-Fold) and on KITTI-360 (58.3 mIoU). Our full pipeline is accessible at https://github.com/drprojects/DeepViewAgg, and only requires raw 3D scans and a set of images and poses.

Despite plenty of efforts focusing on improving the domain adaptation ability (DA) under unsupervised or few-shot semi-supervised settings, recently the solution of active learning started to attract more attention due to its suitability in transferring model in a more practical way with limited annotation resource on target data. Nevertheless, most active learning methods are not inherently designed to handle domain gap between data distribution, on the other hand, some active domain adaptation methods (ADA) usually requires complicated query functions, which is vulnerable to overfitting. In this work, we propose a concise but effective ADA method called Select-by-Distinctive-Margin (SDM), which consists of a maximum margin loss and a margin sampling algorithm for data selection. We provide theoretical analysis to show that SDM works like a Support Vector Machine, storing hard examples around decision boundaries and exploiting them to find informative and transferable data. In addition, we propose two variants of our method, one is designed to adaptively adjust the gradient from margin loss, the other boosts the selectivity of margin sampling by taking the gradient direction into account. We benchmark SDM with standard active learning setting, demonstrating our algorithm achieves competitive results with good data scalability.

Representational learning forms the backbone of most deep learning applications, and the value of a learned representation is intimately tied to its information content regarding different factors of variation. Finding good representations depends on the nature of supervision and the learning algorithm. We propose a novel algorithm that utilizes a weak form of supervision where the data is partitioned into sets according to certain inactive (common) factors of variation which are invariant across elements of each set. Our key insight is that by seeking correspondence between elements of different sets, we learn strong representations that exclude the inactive factors of variation and isolate the active factors that vary within all sets. As a consequence of focusing on the active factors, our method can leverage a mix of set-supervised and wholly unsupervised data, which can even belong to a different domain. We tackle the challenging problem of synthetic-to-real object pose transfer, without pose annotations on anything, by isolating pose information which generalizes to the category level and across the synthetic/real domain gap. The method can also boost performance in supervised settings, by strengthening intermediate representations, as well as operate in practically attainable scenarios with set-supervised natural images, where quantity is limited and nuisance factors of variation are more plentiful.

Vision Transformers (ViT) serve as powerful vision models. Unlike convolutional neural networks, which dominated vision research in previous years, vision transformers enjoy the ability to capture long-range dependencies in the data. Nonetheless, an integral part of any transformer architecture, the self-attention mechanism, suffers from high latency and inefficient memory utilization, making it less suitable for high-resolution input images. To alleviate these shortcomings, hierarchical vision models locally employ self-attention on non-interleaving windows. This relaxation reduces the complexity to be linear in the input size; however, it limits the cross-window interaction, hurting the model performance. In this paper, we propose a new shift-invariant local attention layer, called query and attend (QnA), that aggregates the input locally in an overlapping manner, much like convolutions. The key idea behind QnA is to introduce learned queries, which allow fast and efficient implementation. We verify the effectiveness of our layer by incorporating it into a hierarchical vision transformer model. We show improvements in speed and memory complexity while achieving comparable accuracy with state-of-the-art models. Finally, our layer scales especially well with window size, requiring up to x10 less memory while being up to x5 faster than existing methods. The code is publicly available at https://github.com/moabarar/qna.

We propose a novel multimodal architecture for Scene Text Visual Question Answering (STVQA), named Layout-Aware Transformer (LaTr). The task of STVQA requires models to reason over different modalities. Thus, we first investigate the impact of each modality, and reveal the importance of the language module, especially when enriched with layout information. Accounting for this, we propose a single objective pre-training scheme that requires only text and spatial cues. We show that applying this pre-training scheme on scanned documents has certain advantages over using natural images, despite the domain gap. Scanned documents are easy to procure, text-dense and have a variety of layouts, helping the model learn various spatial cues (e.g. left-of, below etc.) by tying together language and layout information. Compared to existing approaches, our method performs vocabulary-free decoding and, as shown, generalizes well beyond the training vocabulary. We further demonstrate that LaTr improves robustness towards OCR errors, a common reason for failure cases in STVQA. In addition, by leveraging a vision transformer, we eliminate the need for an external object detector. LaTr outperforms state-of-the-art STVQA methods on multiple datasets. In particular, +7.6% on TextVQA, +10.8% on ST-VQA and +4.0% on OCR-VQA (all absolute accuracy numbers).

We present LASER, an image-based Monte Carlo Localization (MCL) framework for 2D floor maps. LASER introduces the concept of latent space rendering, where 2D pose hypotheses on the floor map are directly rendered into a geometrically-structured latent space by aggregating viewing ray features. Through a tightly coupled rendering codebook scheme, the viewing ray features are dynamically determined at rendering-time based on their geometries (i.e. length, incident-angle), endowing our representation with view-dependent fine-grain variability. Our codebook scheme effectively disentangles feature encoding from rendering, allowing the latent space rendering to run at speeds above 10KHz. Moreover, through metric learning, our geometrically-structured latent space is common to both pose hypotheses and query images with arbitrary field of views. As a result, LASER achieves state-of-the-art performance on large-scale indoor localization datasets (i.e. ZInD and Structured3D) for both panorama and perspective image queries, while significantly outperforming existing learning-based methods in speed.

Adversarial training (AT) is always formulated as a minimax problem, of which the performance depends on the inner optimization that involves the generation of adversarial examples (AEs). Most previous methods adopt Projected Gradient Decent (PGD) with manually specifying attack parameters for AE generation. A combination of the attack parameters can be referred to as an attack strategy. Several works have revealed that using a fixed attack strategy to generate AEs during the whole training phase limits the model robustness and propose to exploit different attack strategies at different training stages to improve robustness. But those multi-stage hand-crafted attack strategies need much domain expertise, and the robustness improvement is limited. In this paper, we propose a novel framework for adversarial training by introducing the concept of "learnable attack strategy", dubbed LAS-AT, which learns to automatically produce attack strategies to improve the model robustness. Our framework is composed of a target network that uses AEs for training to improve robustness, and a strategy network that produces attack strategies to control the AE generation. Experimental evaluations on three benchmark databases demonstrate the superiority of the proposed method, and the proposed method outperforms state-of-the-art adversarial training methods.

Far beyond learning long-range interactions of natural language, transformers are becoming the de-facto standard for many vision tasks with their power and scalability. Especially with cross-modal tasks between image and text, vector quantized variational autoencoders (VQ-VAEs) are widely used to make a raw RGB image into a sequence of feature vectors. To better leverage the correlation between image and text, we propose L-Verse, a novel architecture consisting of feature-augmented variational autoencoder (AugVAE) and bidirectional auto-regressive transformer (BiART) for image-to-text and text-to-image generation. Our AugVAE shows the state-of-the-art reconstruction performance on ImageNet1K validation set, along with the robustness to unseen images in the wild. Unlike other models, BiART can distinguish between image (or text) as a conditional reference and a generation target. L-Verse can be directly used for image-to-text or text-to-image generation without any finetuning or extra object detection framework. In quantitative and qualitative experiments, L-Verse shows impressive results against previous methods in both image-to-text and text-to-image generation on MS-COCO Captions. We furthermore assess the scalability of L-Verse architecture on Conceptual Captions and present the initial result of bidirectional vision-language representation learning on general domain.

There is a growing discrepancy in computer vision between large-scale models that achieve state-of-the-art performance and models that are affordable in practical applications. In this paper we address this issue and significantly bridge the gap between these two types of models. Throughout our empirical investigation we do not aim to necessarily propose a new method, but strive to identify a robust and effective recipe for making state-of-the-art large scale models affordable in practice. We demonstrate that, when performed correctly, knowledge distillation can be a powerful tool for reducing the size of large models without compromising their performance. In particular, we uncover that there are certain implicit design choices, which may drastically affect the effectiveness of distillation. Our key contribution is the explicit identification of these design choices, which were not previously articulated in the literature. We back up our findings by a comprehensive empirical study, demonstrate compelling results on a wide range of vision datasets and, in particular, obtain a state-of-the-art ResNet-50 model for ImageNet, which achieves 82.8% top-1 accuracy.

Knowledge distillation becomes a de facto standard to improve the performance of small neural networks. Most of the previous works propose to regress the representational features from the teacher to the student in a one-to-one spatial matching fashion. However, people tend to overlook the fact that, due to the architecture differences, the semantic information on the same spatial location usually vary. This greatly undermines the underlying assumption of the one-to-one distillation approach. To this end, we propose a novel one-to-all spatial matching knowledge distillation approach. Specifically, we allow each pixel of the teacher feature to be distilled to all spatial locations of the student features given its similarity, which is generated from a target-aware transformer. Our approach surpasses the state-of-the-art methods by a significant margin on various computer vision benchmarks, such as ImageNet, Pascal VOC and COCOStuff10k. Code is available at https://github.com/sihaoevery/TaT.

We propose a robust and accurate method for estimating the 3D poses of two hands in close interaction from a single color image. This is a very challenging problem, as large occlusions and many confusions between the joints may happen. State-of-the-art methods solve this problem by regressing a heatmap for each joint, which requires solving two problems simultaneously: localizing the joints and recognizing them. In this work, we propose to separate these tasks by relying on a CNN to first localize joints as 2D keypoints, and on self-attention between the CNN features at these keypoints to associate them with the corresponding hand joint. The resulting architecture, which we call "Keypoint Transformer", is highly efficient as it achieves state-of-the-art performance with roughly half the number of model parameters on the InterHand2.6M dataset. We also show it can be easily extended to estimate the 3D pose of an object manipulated by one or two hands with high performance. Moreover, we created a new dataset of more than 75,000 images of two hands manipulating an object fully annotated in 3D and will make it publicly available.

We consider the problem of reconstructing the depth of dynamic objects from videos. Recent progress in dynamic video depth prediction has focused on improving the output of monocular depth estimators by means of multi-view constraints while imposing little to no restrictions on the deformation of the dynamic parts of the scene. However, the theory of Non-Rigid Structure from Motion prescribes to constrain the deformations for 3D reconstruction. We thus propose a new model that departs significantly from this prior work. The idea is to fit a dynamic point cloud to the video data using Sinkhorn's algorithm to associate the 3D points to 2D pixels and use a differentiable point renderer to ensure the compatibility of the 3D deformations with the measured optical flow. In this manner, our algorithm, called Keypoint Transporter, models the overall deformation of the object within the entire video, so it can constrain the reconstruction correspondingly. Compared to weaker deformation models, this significantly reduces the reconstruction ambiguity and, for dynamic objects, allows Keypoint Transporter to obtain reconstructions of the quality superior or at least comparable to prior approaches while being much faster and reliant on a pre-trained monocular depth estimator network. To assess the method, we evaluate on new datasets of synthetic videos depicting dynamic humans and animals with ground-truth depth. We also show qualitative results on crowd-sourced real-world videos of pets.

Video summarization has recently engaged increasing attention in computer vision communities. However, the scarcity of annotated data has been a key obstacle in this task. To address it, this work explores a new solution for video summarization by transferring samples from a correlated task (i.e., video moment localization) equipped with abundant training data. Our main insight is that the annotated video moments also indicate the semantic highlights of a video, essentially similar to video summary. Approximately, the video summary can be treated as a sparse, redundancy-free version of the video moments. Inspired by this observation, we propose an importance Propagation based collaborative Teaching Network (iPTNet). It consists of two separate modules that conduct video summarization and moment localization, respectively. Each module estimates a frame-wise importance map for indicating keyframes or moments. To perform cross-task sample transfer, we devise an importance propagation module that realizes the conversion between summarization-guided and localization-guided importance maps. This way critically enables optimizing one of the tasks using the data from the other task. Additionally, in order to avoid error amplification caused by batch-wise joint training, we devise a collaborative teaching scheme, which adopts a cross-task mean teaching strategy to realize the joint optimization of the two tasks and provide robust frame-level teaching signals. Extensive experiments on video summarization benchmarks demonstrate that iPTNet significantly outperforms previous state-of-the-art video summarization methods, serving as an effective solution that overcomes the data scarcity issue in video summarization.