CVPR 2023 论文列表

3D semantic scene graphs are a powerful holistic representation as they describe the individual objects and depict the relation between them. They are compact high-level graphs that enable many tasks requiring scene reasoning. In real-world settings, existing 3D estimation methods produce robust predictions that mostly rely on dense inputs. In this work, we propose a real-time framework that incrementally builds a consistent 3D semantic scene graph of a scene given an RGB image sequence. Our method consists of a novel incremental entity estimation pipeline and a scene graph prediction network. The proposed pipeline simultaneously reconstructs a sparse point map and fuses entity estimation from the input images. The proposed network estimates 3D semantic scene graphs with iterative message passing using multi-view and geometric features extracted from the scene entities. Extensive experiments on the 3RScan dataset show the effectiveness of the proposed method in this challenging task, outperforming state-of-the-art approaches.

We propose a method for in-hand 3D scanning of an unknown object with a monocular camera. Our method relies on a neural implicit surface representation that captures both the geometry and the appearance of the object, however, by contrast with most NeRF-based methods, we do not assume that the camera-object relative poses are known. Instead, we simultaneously optimize both the object shape and the pose trajectory. As direct optimization over all shape and pose parameters is prone to fail without coarse-level initialization, we propose an incremental approach that starts by splitting the sequence into carefully selected overlapping segments within which the optimization is likely to succeed. We reconstruct the object shape and track its poses independently within each segment, then merge all the segments before performing a global optimization. We show that our method is able to reconstruct the shape and color of both textured and challenging texture-less objects, outperforms classical methods that rely only on appearance features, and that its performance is close to recent methods that assume known camera poses.

Deep neural networks have achieved unprecedented success on diverse vision tasks. However, they are vulnerable to adversarial noise that is imperceptible to humans. This phenomenon negatively affects their deployment in real-world scenarios, especially security-related ones. To evaluate the robustness of a target model in practice, transfer-based attacks craft adversarial samples with a local model and have attracted increasing attention from researchers due to their high efficiency. The state-of-the-art transfer-based attacks are generally based on data augmentation, which typically augments multiple training images from a linear path when learning adversarial samples. However, such methods selected the image augmentation path heuristically and may augment images that are semantics-inconsistent with the target images, which harms the transferability of the generated adversarial samples. To overcome the pitfall, we propose the Path-Augmented Method (PAM). Specifically, PAM first constructs a candidate augmentation path pool. It then settles the employed augmentation paths during adversarial sample generation with greedy search. Furthermore, to avoid augmenting semantics-inconsistent images, we train a Semantics Predictor (SP) to constrain the length of the augmentation path. Extensive experiments confirm that PAM can achieve an improvement of over 4.8% on average compared with the state-of-the-art baselines in terms of the attack success rates.

Multi-modal image-text models such as CLIP and LiT have demonstrated impressive performance on image classification benchmarks and their zero-shot generalization ability is particularly exciting. While the top-5 zero-shot accuracies of these models are very high, the top-1 accuracies are much lower (over 25% gap in some cases). We investigate the reasons for this performance gap and find that many of the failure cases are caused by ambiguity in the text prompts. First, we develop a simple and efficient zero-shot post-hoc method to identify images whose top-1 prediction is likely to be incorrect, by measuring consistency of the predictions w.r.t. multiple prompts and image transformations. We show that our procedure better predicts mistakes, outperforming the popular max logit baseline on selective prediction tasks. Next, we propose a simple and efficient way to improve accuracy on such uncertain images by making use of the WordNet hierarchy; specifically we augment the original class by incorporating its parent and children from the semantic label hierarchy, and plug the augmentation into text prompts. We conduct experiments on both CLIP and LiT models with five different ImageNet- based datasets. For CLIP, our method improves the top-1 accuracy by 17.13% on the uncertain subset and 3.6% on the entire ImageNet validation set. We also show that our method improves across ImageNet shifted datasets, four other datasets, and other model architectures such as LiT. Our proposed method is hyperparameter-free, requires no additional model training and can be easily scaled to other large multi-modal architectures. Code is available at https://github.com/gyhandy/Hierarchy-CLIP.

The task of weakly supervised temporal action localization targets at generating temporal boundaries for actions of interest, meanwhile the action category should also be classified. Pseudo-label-based methods, which serve as an effective solution, have been widely studied recently. However, existing methods generate pseudo labels during training and make predictions during testing under different pipelines or settings, resulting in a gap between training and testing. In this paper, we propose to generate high-quality pseudo labels from the predicted action boundaries. Nevertheless, we note that existing post-processing, like NMS, would lead to information loss, which is insufficient to generate high-quality action boundaries. More importantly, transforming action boundaries into pseudo labels is quite challenging, since the predicted action instances are generally overlapped and have different confidence scores. Besides, the generated pseudo-labels can be fluctuating and inaccurate at the early stage of training. It might repeatedly strengthen the false predictions if there is no mechanism to conduct self-correction. To tackle these issues, we come up with an effective pipeline for learning better pseudo labels. Firstly, we propose a Gaussian weighted fusion module to preserve information of action instances and obtain high-quality action boundaries. Second, we formulate the pseudo-label generation as an optimization problem under the constraints in terms of the confidence scores of action instances. Finally, we introduce the idea of Delta pseudo labels, which enables the model with the ability of self-correction. Our method achieves superior performance to existing methods on two benchmarks, THUMOS14 and ActivityNet1.3, achieving gains of 1.9% on THUMOS14 and 3.7% on ActivityNet1.3 in terms of average mAP.

We present an extension to masked autoencoders (MAE) which improves on the representations learnt by the model by explicitly encouraging the learning of higher scene-level features. We do this by: (i) the introduction of a perceptual similarity term between generated and real images (ii) incorporating several techniques from the adversarial training literature including multi-scale training and adaptive discriminator augmentation. The combination of these results in not only better pixel reconstruction but also representations which appear to capture better higher-level details within images. More consequentially, we show how our method, Perceptual MAE, leads to better performance when used for downstream tasks outperforming previous methods. We achieve 78.1% top-1 accuracy linear probing on ImageNet-1K and up to 88.1% when fine-tuning, with similar results for other downstream tasks, all without use of additional pre-trained models or data.

We propose a margin-based loss for tuning joint vision-language models so that their gradient-based explanations are consistent with region-level annotations provided by humans for relatively smaller grounding datasets. We refer to this objective as Attention Mask Consistency (AMC) and demonstrate that it produces superior visual grounding results than previous methods that rely on using vision-language models to score the outputs of object detectors. Particularly, a model trained with AMC on top of standard vision-language modeling objectives obtains a state-of-the-art accuracy of 86.49% in the Flickr30k visual grounding benchmark, an absolute improvement of 5.38% when compared to the best previous model trained under the same level of supervision. Our approach also performs exceedingly well on established benchmarks for referring expression comprehension where it obtains 80.34% accuracy in the easy test of RefCOCO+, and 64.55% in the difficult split. AMC is effective, easy to implement, and is general as it can be adopted by any vision-language model, and can use any type of region annotations.

Vision-and-Language Navigation (VLN) is the task that requires an agent to navigate through the environment based on natural language instructions. At each step, the agent takes the next action by selecting from a set of navigable locations. In this paper, we aim to take one step further and explore whether the agent can benefit from generating the potential future view during navigation. Intuitively, humans will have an expectation of how the future environment will look like, based on the natural language instructions and surrounding views, which will aid correct navigation. Hence, to equip the agent with this ability to generate the semantics of future navigation views, we first propose three proxy tasks during the agent's in-domain pre-training: Masked Panorama Modeling (MPM), Masked Trajectory Modeling (MTM), and Action Prediction with Image Generation (APIG). These three objectives teach the model to predict missing views in a panorama (MPM), predict missing steps in the full trajectory (MTM), and generate the next view based on the full instruction and navigation history (APIG), respectively. We then fine-tune the agent on the VLN task with an auxiliary loss that minimizes the difference between the view semantics generated by the agent and the ground truth view semantics of the next step. Empirically, our VLN-SIG achieves the new state-of-the-art on both the Room-to-Room dataset and the CVDN dataset. We further show that our agent learns to fill in missing patches in future views qualitatively, which brings more interpretability over agents' predicted actions. Lastly, we demonstrate that learning to predict future view semantics also enables the agent to have better performance on longer paths.

Table structure recognition aims to extract the logical and physical structure of unstructured table images into a machine-readable format. The latest end-to-end image-to-text approaches simultaneously predict the two structures by two decoders, where the prediction of the physical structure (the bounding boxes of the cells) is based on the representation of the logical structure. However, as the logical representation lacks the local visual information, the previous methods often produce imprecise bounding boxes. To address this issue, we propose an end-to-end sequential modeling framework for table structure recognition called VAST. It contains a novel coordinate sequence decoder triggered by the representation of the non-empty cell from the logical structure decoder. In the coordinate sequence decoder, we model the bounding box coordinates as a language sequence, where the left, top, right and bottom coordinates are decoded sequentially to leverage the inter-coordinate dependency. Furthermore, we propose an auxiliary visual-alignment loss to enforce the logical representation of the non-empty cells to contain more local visual details, which helps produce better cell bounding boxes. Extensive experiments demonstrate that our proposed method can achieve state-of-the-art results in both logical and physical structure recognition. The ablation study also validates that the proposed coordinate sequence decoder and the visual-alignment loss are the keys to the success of our method.

Despite advances in Visual Question Answering (VQA), the ability of models to assess their own correctness remains underexplored. Recent work has shown that VQA models, out-of-the-box, can have difficulties abstaining from answering when they are wrong. The option to abstain, also called Selective Prediction, is highly relevant when deploying systems to users who must trust the system's output (e.g., VQA assistants for users with visual impairments). For such scenarios, abstention can be especially important as users may provide out-of-distribution (OOD) or adversarial inputs that make incorrect answers more likely. In this work, we explore Selective VQA in both in-distribution (ID) and OOD scenarios, where models are presented with mixtures of ID and OOD data. The goal is to maximize the number of questions answered while minimizing the risk of error on those questions. We propose a simple yet effective Learning from Your Peers (LYP) approach for training multimodal selection functions for making abstention decisions. Our approach uses predictions from models trained on distinct subsets of the training data as targets for optimizing a Selective VQA model. It does not require additional manual labels or held-out data and provides a signal for identifying examples that are easy/difficult to generalize to. In our extensive evaluations, we show this benefits a number of models across different architectures and scales. Overall, for ID, we reach 32.92% in the selective prediction metric coverage at 1% risk of error (C@1%) which doubles the previous best coverage of 15.79% on this task. For mixed ID/OOD, using models' softmax confidences for abstention decisions performs very poorly, answering <5% of questions at 1% risk of error even when faced with only 10% OOD examples, but a learned selection function with LYP can increase that to 25.38% C@1%.

Despite their success, vision transformers still remain vulnerable to image corruptions, such as noise or blur. Indeed, we find that the vulnerability mainly stems from the unstable self-attention mechanism, which is inherently built upon patch-based inputs and often becomes overly sensitive to the corruptions across patches. For example, when we only occlude a small number of patches with random noise (e.g., 10%), these patch corruptions would lead to severe accuracy drops and greatly distract intermediate attention layers. To address this, we propose a new training method that improves the robustness of transformers from a new perspective -- reducing sensitivity to patch corruptions (RSPC). Specifically, we first identify and occlude/corrupt the most vulnerable patches and then explicitly reduce sensitivity to them by aligning the intermediate features between clean and corrupted examples. We highlight that the construction of patch corruptions is learned adversarially to the following feature alignment process, which is particularly effective and essentially different from existing methods. In experiments, our RSPC greatly improves the stability of attention layers and consistently yields better robustness on various benchmarks, including CIFAR-10/100-C, ImageNet-A, ImageNet-C, and ImageNet-P.

Semantic segmentation involves classifying each pixel into one of a pre-defined set of object/stuff classes. Such a fine-grained detection and localization of objects in the scene is challenging by itself. The complexity increases manifold in the presence of blur. With cameras becoming increasingly light-weight and compact, blur caused by motion during capture time has become unavoidable. Most research has focused on improving segmentation performance for sharp clean images and the few works that deal with degradations, consider motion-blur as one of many generic degradations. In this work, we focus exclusively on motion-blur and attempt to achieve robustness for semantic segmentation in its presence. Based on the observation that segmentation annotations can be used to generate synthetic space-variant blur, we propose a Class-Centric Motion-Blur Augmentation (CCMBA) strategy. Our approach involves randomly selecting a subset of semantic classes present in the image and using the segmentation map annotations to blur only the corresponding regions. This enables the network to simultaneously learn semantic segmentation for clean images, images with egomotion blur, as well as images with dynamic scene blur. We demonstrate the effectiveness of our approach for both CNN and Vision Transformer-based semantic segmentation networks on PASCAL VOC and Cityscapes datasets. We also illustrate the improved generalizability of our method to complex real-world blur by evaluating on the commonly used deblurring datasets GoPro and REDS.

Recent research in robust optimization has shown an overfitting-like phenomenon in which models trained against adversarial attacks exhibit higher robustness on the training set compared to the test set. Although previous work provided theoretical explanations for this phenomenon using a robust PAC-Bayesian bound over the adversarial test error, related algorithmic derivations are at best only loosely connected to this bound, which implies that there is still a gap between their empirical success and our understanding of adversarial robustness theory. To close this gap, in this paper we consider a different form of the robust PAC-Bayesian bound and directly minimize it with respect to the model posterior. The derivation of the optimal solution connects PAC-Bayesian learning to the geometry of the robust loss surface through a Trace of Hessian (TrH) regularizer that measures the surface flatness. In practice, we restrict the TrH regularizer to the top layer only, which results in an analytical solution to the bound whose computational cost does not depend on the network depth. Finally, we evaluate our TrH regularization approach over CIFAR-10/100 and ImageNet using Vision Transformers (ViT) and compare against baseline adversarial robustness algorithms. Experimental results show that TrH regularization leads to improved ViT robustness that either matches or surpasses previous state-of-the-art approaches while at the same time requires less memory and computational cost.

Retrieval augmented models are becoming increasingly popular for computer vision tasks after their recent success in NLP problems. The goal is to enhance the recognition capabilities of the model by retrieving similar examples for the visual input from an external memory set. In this work, we introduce an attention-based memory module, which learns the importance of each retrieved example from the memory. Compared to existing approaches, our method removes the influence of the irrelevant retrieved examples, and retains those that are beneficial to the input query. We also thoroughly study various ways of constructing the memory dataset. Our experiments show the benefit of using a massive-scale memory dataset of 1B image-text pairs, and demonstrate the performance of different memory representations. We evaluate our method in three different classification tasks, namely long-tailed recognition, learning with noisy labels, and fine-grained classification, and show that it achieves state-of-the-art accuracies in ImageNet-LT, Places-LT and Webvision datasets.

Recently, self-attention networks achieve impressive performance in point cloud segmentation due to their superiority in modeling long-range dependencies. However, compared to self-attention mechanism, we find graph convolutions show a stronger ability in capturing local geometry information with less computational cost. In this paper, we employ a hybrid architecture design to construct our Graph Convolution Network with Attentive Filtering (AF-GCN), which takes advantage of both graph convolution and self-attention mechanism. We adopt graph convolutions to aggregate local features in the shallow encoder stages, while in the deeper stages, we propose a self-attention-like module named Graph Attentive Filter (GAF) to better model long-range contexts from distant neighbors. Besides, to further improve graph representation for point cloud segmentation, we employ a Spatial Feature Projection (SFP) module for graph convolutions which helps to handle spatial variations of unstructured point clouds. Finally, a graph-shared down-sampling and up-sampling strategy is introduced to make full use of the graph structures in point cloud processing. We conduct extensive experiments on multiple datasets including S3DIS, ScanNetV2, Toronto-3D, and ShapeNetPart. Experimental results show our AF-GCN obtains competitive performance.

Despite the broad interest in meta-learning, the generalization problem remains one of the significant challenges in this field. Existing works focus on meta-generalization to unseen tasks at the meta-level by regularizing the meta-loss, while ignoring that adapted models may not generalize to the task domains at the adaptation level. In this paper, we propose a new regularization mechanism for meta-learning -- Minimax-Meta Regularization, which employs inverted regularization at the inner loop and ordinary regularization at the outer loop during training. In particular, the inner inverted regularization makes the adapted model more difficult to generalize to task domains; thus, optimizing the outer-loop loss forces the meta-model to learn meta-knowledge with better generalization. Theoretically, we prove that inverted regularization improves the meta-testing performance by reducing generalization errors. We conduct extensive experiments on the representative scenarios, and the results show that our method consistently improves the performance of meta-learning algorithms.

Domain generalization (DG) tends to alleviate the poor generalization capability of deep neural networks by learning model with multiple source domains. A classical solution to DG is domain augmentation, the common belief of which is that diversifying source domains will be conducive to the out-of-distribution generalization. However, these claims are understood intuitively, rather than mathematically. Our explorations empirically reveal that the correlation between model generalization and the diversity of domains may be not strictly positive, which limits the effectiveness of domain augmentation. This work therefore aim to guarantee and further enhance the validity of this strand. To this end, we propose a new perspective on DG that recasts it as a convex game between domains. We first encourage each diversified domain to enhance model generalization by elaborately designing a regularization term based on supermodularity. Meanwhile, a sample filter is constructed to eliminate low-quality samples, thereby avoiding the impact of potentially harmful information. Our framework presents a new avenue for the formal analysis of DG, heuristic analysis and extensive experiments demonstrate the rationality and effectiveness.

Recent 3D face reconstruction methods have made significant advances in geometry prediction, yet further cosmetic improvements are limited by lagged albedo because inferring albedo from appearance is an ill-posed problem. Although some existing methods consider prior knowledge from illumination to improve albedo estimation, they still produce a light-skin bias due to racially biased albedo models and limited light constraints. In this paper, we reconsider the relationship between albedo and face attributes and propose an ID2Albedo to directly estimate albedo without constraining illumination. Our key insight is that intrinsic semantic attributes such as race, skin color, and age can constrain the albedo map. We first introduce visual-textual cues and design a semantic loss to supervise facial albedo estimation. Specifically, we pre-define text labels such as race, skin color, age, and wrinkles. Then, we employ the text-image model (CLIP) to compute the similarity between the text and the input image, and assign a pseudo-label to each facial image. We constrain generated albedos in the training phase to have the same attributes as the inputs. In addition, we train a high-quality, unbiased facial albedo generator and utilize the semantic loss to learn the mapping from illumination-robust identity features to the albedo latent codes. Finally, our ID2Albedo is trained in a self-supervised way and outperforms state-of-the-art albedo estimation methods in terms of accuracy and fidelity. It is worth mentioning that our approach has excellent generalizability and fairness, especially on in-the-wild data.

Cross-modal retrieval across image and text modalities is a challenging task due to its inherent ambiguity: An image often exhibits various situations, and a caption can be coupled with diverse images. Set-based embedding has been studied as a solution to this problem. It seeks to encode a sample into a set of different embedding vectors that capture different semantics of the sample. In this paper, we present a novel set-based embedding method, which is distinct from previous work in two aspects. First, we present a new similarity function called smooth-Chamfer similarity, which is designed to alleviate the side effects of existing similarity functions for set-based embedding. Second, we propose a novel set prediction module to produce a set of embedding vectors that effectively captures diverse semantics of input by the slot attention mechanism. Our method is evaluated on the COCO and Flickr30K datasets across different visual backbones, where it outperforms existing methods including ones that demand substantially larger computation at inference.