ECCV 2022 论文列表

"Neural networks are vulnerable to adversarial attacks: adding well-crafted, imperceptible perturbations to their input can modify their output. Adversarial training is one of the most effective approaches in training robust models against such attacks. However, it is much slower than vanilla training of neural networks since it needs to construct adversarial examples for the entire training data at every iteration, hampering its effectiveness. Recently, Fast Adversarial Training (FAT) was proposed that can obtain robust models efficiently. However, the reasons behind its success are not fully understood, and more importantly, it can only train robust models for $\ell_\infty$-bounded attacks as it uses FGSM during training. In this paper, by leveraging the theory of coreset selection, we show how selecting a small subset of training data provides a general, more principled approach toward reducing the time complexity of robust training. Unlike existing methods, our approach can be adapted to a wide variety of training objectives, including TRADES, $\ell_p$-PGD, and Perceptual Adversarial Training (PAT). Our experimental results indicate that our approach speeds up adversarial training by 2-3 times while experiencing a slight reduction in the clean and robust accuracy."

"The rise of transformers in vision tasks not only advances network backbone designs, but also starts a brand-new page to achieve end-to-end image recognition (e.g., object detection and panoptic segmentation). Originated from Natural Language Processing (NLP), transformer architectures, consisting of self-attention and cross-attention, effectively learn long-range interactions between elements in a sequence. However, we observe that most existing transformer-based vision models simply borrow the idea from NLP, neglecting the crucial difference between languages and images, particularly the extremely large sequence length of spatially flattened pixel features. This subsequently impedes the learning in cross-attention between pixel features and object queries. In this paper, we rethink the relationship between pixels and object queries and propose to reformulate the cross-attention learning as a clustering process. Inspired by the traditional k-means clustering algorithm, we develop a k-means Mask Xformer (kMaX-DeepLab) for segmentation tasks, which not only improves the state-of-the-art, but also enjoys a simple and elegant design. As a result, our kMaX-DeepLab achieves a new state-of-the-art performance on COCO val set with 58.0% PQ, and Cityscapes val set with 68.4% PQ, 44.0% AP, and 83.5% mIoU without test-time augmentation or external dataset. We hope our work can shed some light on designing transformers tailored for vision tasks. Code and models are available at https://github.com/google-research/deeplab2"

"Novelty detection is a key capability for practical machine learning in the real world, where models operate in non-stationary conditions and are repeatedly exposed to new, unseen data. Yet, most current novelty detection approaches have been developed exclusively for static, offline use. They scale poorly under more realistic, continual learning regimes in which data distribution shifts occur. To address this critical gap, this paper proposes incDFM (incremental Deep Feature Modeling), a self-supervised continual novelty detector. The method builds a statistical model over the space of intermediate features produced by a deep network, and utilizes feature reconstruction errors as uncertainty scores to guide the detection of novel samples. Most importantly, incDFM estimates the statistical model incrementally (via several iterations within a task), instead of a single-shot. Each time it selects only the most confident novel samples which will then guide subsequent recruitment incrementally. For a certain task where the ML model encounters a mixture of old and novel data, the detector flags novel samples to incorporate them to old knowledge. Then the detector is updated with the flagged novel samples, in preparation for a next task. To quantify and benchmark performance, we adapted multiple datasets for continual learning: CIFAR-10, CIFAR-100, SVHN, iNaturalist, and the 8-dataset. Our experiments show that incDFM achieves state of the art continual novelty detection performance. Furthermore, when examined in the greater context of continual learning for classification, our method is successful in minimizing catastrophic forgetting and error propagation."

"Standard spatial convolutions assume input data with a regular neighborhood structure. Existing methods typically generalize convolution to the irregular point cloud domain by fixing a regular ""view"" through e.g. a fixed neighborhood size, where the convolution kernel size remains the same for each point. However, since point clouds are not as structured as images, the fixed neighbor number gives an unfortunate inductive bias. We present a novel graph convolution named Difference Graph Convolution (diffConv), which does not rely on a regular view. diffConv operates on spatially-varying and density-dilated neighborhoods, which are further adapted by a learned masked attention mechanism. Experiments show that our model is very robust to the noise, obtaining state-of-the-art performance in 3D shape classification and scene understanding tasks, along with a faster inference speed."

"Existing temporal action detection (TAD) methods rely on large training data including segment-level annotations, limited to recognizing previously seen classes alone during inference. Collecting and annotating a large training set for each class of interest is costly and hence unscalable. Zero-shot TAD (ZS-TAD) resolves this obstacle by enabling a pre-trained model to recognize any unseen action classes. Meanwhile, ZS-TAD is also much more challenging with significantly less investigation. Inspired by the success of zero-shot image classification aided by vision-language (ViL) models such as CLIP, we aim to tackle the more complex TAD task. An intuitive method is to integrate an off-the-shelf proposal detector with CLIP style classification. However, due to the sequential localization (e.g, proposal generation) and classification design, it is prone to localization error propagation. To overcome this problem, in this paper we propose a novel zero-Shot Temporal Action detection model via vision-LanguagE prompting (STALE). Such a novel design effectively eliminates the dependence between localization and classification by breaking the route for error propagation in-between. We further introduce an interaction mechanism between classification and localization for improved optimization. Extensive experiments on standard ZS-TAD video benchmarks show that our STALE significantly outperforms state-of-the-art alternatives. Besides, our model also yields superior results on supervised TAD over recent strong competitors. The PyTorch implementation of STALE is available on https://github.com/sauradip/STALE."

"The existing methods for reflection removal mainly focus on removing blurry and weak reflection artifacts and thus often fail to work with severe and strong reflection artifacts. However, in many cases, real reflection artifacts are sharp and intensive enough such that even humans cannot completely distinguish between the transmitted and reflected scenes. In this paper, we attempt to remove such challenging reflection artifacts using 360-degree images. We adopt the zero-shot learning scheme to avoid the burden of collecting paired data for supervised learning and the domain gap between different datasets. We first search for the reference image of the reflected scene in a 360-degree image based on the reflection geometry, which is then used to guide the network to restore the faithful colors of the reflection image. We collect 30 test 360-degree images exhibiting challenging reflection artifacts and demonstrate that the proposed method outperforms the existing state-of-the-art methods on 360-degree images."

"Object pose estimation is an important component of most vision pipelines for embodied agents, as well as in 3D vision more generally. In this paper we tackle the problem of estimating the pose of novel object categories in a zero-shot manner. This extends much of the existing literature by removing the need for pose-labelled datasets or category-specific CAD models for training or inference. Specifically, we make the following contributions. First, we formalise the zero-shot, category-level pose estimation problem and frame it in a way that is most applicable to real-world embodied agents. Secondly, we propose a novel method based on semantic correspondences from a self-supervised vision transformer to solve the pose estimation problem. We further re-purpose the recent CO3D dataset to present a controlled and realistic test setting. Finally, we demonstrate that all baselines for our proposed task perform poorly, and show that our method provides a six-fold improvement in average rotation accuracy at 30 degrees. Our code is available at https://github.com/applied-ai-lab/zero-shot-pose."

"Zero-shot fine-grained recognition is an important classification task, whose goal is to recognize visually very similar classes, including the ones without training images. Despite recent advances on the development of zero-shot fine-grained recognition methods, the robustness of such models to adversarial attacks is not well understood. On the other hand, adversarial attacks have been widely studied for conventional classification with visually distinct classes. Such attacks, in particular, universal perturbations that are class-agnostic and ideally should generalize to unseen classes, however, cannot leverage or capture small distinctions among fine-grained classes. Therefore, we propose a compositional attribute-based framework for generating adversarial attacks on zero-shot fine-grained recognition models. To generate attacks that capture small differences between fine-grained classes, generalize well to previously unseen classes and can be applied in real-time, we propose to learn and compose multiple attribute-based universal perturbations (AUPs). Each AUP corresponds to an image-agnostic perturbation on a specific attribute. To build our attack, we compose AUPs with weights obtained by learning a class-attribute compatibility function. To learn the AUPs and the parameters of our model, we minimize a loss, consisting of a ranking loss and a novel utility loss, which ensures AUPs are effectively learned and utilized. By extensive experiments on three datasets for zero-shot fine-grained recognition, we show that our attacks outperform conventional universal classification attacks and transfer well between different recognition architectures."

"Feature pyramid network (FPN) is one of the key components for object detectors. However, there is a long-standing puzzle for researchers that the detection performance of large-scale objects are usually suppressed after introducing FPN. To this end, this paper first revisits FPN in the detection framework and reveals the nature of the success of FPN from the perspective of optimization. Then, we point out that the degraded performance of large-scale objects is due to the arising of improper back-propagation paths after integrating FPN. It makes each level of the backbone network only has the ability to look at the objects within a certain scale range. Based on these analysis, two feasible strategies are proposed to enable each level of the backbone to look at all objects in the FPN-based detection frameworks. Specifically, one is to introduce auxiliary objective functions to make each backbone level directly receive the back-propagation signals of various-scale objects during training. The other is to construct the feature pyramid in a more reasonable way to avoid the irrational back-propagation paths. Extensive experiments on the COCO benchmark validate the soundness of our analysis and the effectiveness of our methods. Without bells and whistles, we demonstrate that our method achieves solid improvements (more than 2%) on various detection frameworks: one-stage, two-stage, anchor-based, anchor-free and transformer-based detectors."

"Stochastic rounding is a critical technique used in low-precision deep neural networks (DNNs) training to ensure good model accuracy. However, it requires a large number of random numbers generated on the fly. This is not a trivial task on the hardware platforms such as FPGA and ASIC. The widely used solution is to introduce random number generators with extra hardware costs. In this paper, we innovatively propose to employ the stochastic property of DNN training process itself and directly extract random numbers from DNNs in a self-sufficient manner. We propose different methods to obtain random numbers from different sources in neural networks and a generator-free framework is proposed for low-precision DNN training on a variety of deep learning tasks. Moreover, we evaluate the quality of the extracted random numbers and find that high-quality random numbers widely exist in DNNs, while their quality can even pass the NIST test suite."

"We present XMem, a video object segmentation architecture for long videos with unified feature memory stores inspired by the Atkinson-Shiffrin memory model. Prior work on video object segmentation typically only uses one type of feature memory. For videos longer than a minute, a single feature memory model tightly links memory consumption and accuracy. In contrast, following the Atkinson-Shiffrin model, we develop an architecture that incorporates multiple independent yet deeply-connected feature memory stores: a rapidly updated sensory memory, a high-resolution working memory, and a compact thus sustained long-term memory. Crucially, we develop a memory potentiation algorithm that routinely consolidates actively used working memory elements into the long-term memory, which avoids memory explosion and minimizes performance decay for long-term prediction. Combined with a new memory reading mechanism, XMem greatly exceeds state-of-the-art performance on long-video datasets while being on par with state-of-the-art methods (that do not work on long videos) on short-video datasets. Code is available at https://hkchengrex.github.io/XMem"

"In computer vision, pre-training models based on large-scale supervised learning have been proven effective over the past few years. However, existing works mostly focus on learning from the individual tasks with the single data source e.g., ImageNet for classification or COCO for detection). This restricted form limits their generalizability and usability due to the lack of vast semantic information from various tasks and data sources. Here, we demonstrate that jointly learning from heterogeneous tasks and multiple data sources contributes to universal visual representation, leading to better transferring results of various downstream tasks. Thus, learning how to bridge the gaps among different tasks and data sources is the key, but it still remains an open question. In this work, we propose a representation learning framework called X-Learner, which learns the universal feature of multiple vision tasks supervised by various sources, with expansion and squeeze stage:1) Expansion Stage: X-Learner learns the task-specific feature to alleviate task interference and enrich the representation by reconciliation layer. 2) Squeeze Stage: X-Learner condenses the model to a reasonable size and learns the universal and generalizable representation for various tasks transferring. Extensive experiments demonstrate that X-Learner achieves strong performance on different tasks without extra annotations, modalities and computational costs compared to existing representation learning methods. Notably, a single X-Learner model shows remarkable gains of 3.0%, 3.3% and 1.8% over current pre-trained models on 12 downstream datasets for classification, object detection and semantic segmentation."

"In this paper, we study the challenging instance-wise vision-language tasks, where the free-form language is required to align with the objects instead of the whole image. To address these tasks, we propose X-DETR, whose architecture has three major components: an object detector, a language encoder, and vision-language alignment. The vision and language streams are independent until the end and they are aligned using an efficient dot-product operation. The whole network is trained end-to-end, such that the detector is optimized for the vision-language tasks instead of an off-the-shelf component. To overcome the limited size of paired object-language annotations, we leverage other weak types of supervision to expand the knowledge coverage. This simple yet effective architecture of X-DETR shows good accuracy and fast speeds for multiple instance-wise vision-language tasks, e.g., 16.4 AP on LVIS detection of 1.2K categories at 20 frames per second without using any LVIS annotation during training. The code is available at https://github.com/amazon-research/cross-modal-detr."

"Few-shot recognition learns a recognition model with very few (e.g., 1 or 5) images per category, and current few-shot learning methods focus on improving the average accuracy over many episodes. We argue that in real-world applications we may often only try one episode instead of many, and hence maximizing the worst-case accuracy is more important than maximizing the average accuracy. We empirically show that a high average accuracy not necessarily means a high worst-case accuracy. Since this objective is not accessible, we propose to reduce the standard deviation and increase the average accuracy simultaneously. In turn, we devise two strategies from the bias-variance tradeoff perspective to implicitly reach this goal: a simple yet effective stability regularization (SR) loss together with model ensemble to reduce variance during fine-tuning, and an adaptability calibration mechanism to reduce the bias. Extensive experiments on benchmark datasets demonstrate the effectiveness of the proposed strategies, which outperforms current state-of-the-art methods with a significant margin in terms of not only average, but also worst-case accuracy."

"Story visualization aims to generate a sequence of images to narrate each sentence in a multi-sentence story with a global consistency across dynamic scenes and characters. Current works still struggle with output images’ quality and consistency, and rely on additional semantic information or auxiliary captioning networks. To address these challenges, we first introduce a new sentence representation, which incorporates word information from all story sentences to mitigate the inconsistency problem. Then, we propose a new discriminator with fusion features and further extend the spatial attention to improve image quality and story consistency. Extensive experiments on different datasets and human evaluation demonstrate the superior performance of our approach, compared to state-of-the-art methods, neither using segmentation masks nor auxiliary captioning networks."

"Object categories are often grouped into a multi-granularity taxonomic hierarchy. Classifying objects at coarser-grained hierarchy requires global and common characteristics, while finer-grained hierarchy classification relies on local and discriminative features. Therefore, humans should also subconsciously focus on different object regions when classifying different hierarchies. This granularity-wise attention is confirmed by our collected human real-time gaze data on different hierarchy classifications. To leverage this mechanism, we propose a Cross-Hierarchical Region Feature (CHRF) learning framework. Specifically, we first design a region feature mining module that imitates humans to learn different granularity-wise attention regions with multi-grained classification tasks. To explore how human attention shifts from one hierarchy to another, we further present a cross-hierarchical orthogonal fusion module to enhance the region feature representation by blending the original feature and an orthogonal component extracted from adjacent hierarchies. Experiments on five hierarchical fine-grained datasets demonstrate the effectiveness of CHRF compared with the state-of-the-art methods. Ablation study and visualization results also consistently verify the advantages of our human attention-oriented modules. The code and dataset are available at https://github.com/visiondom/CHRF."

"Predicting the geographic location (geo-localization) from a single ground-level RGB image taken anywhere in the world is a very challenging problem. The challenges include huge diversity of images due to different environmental scenarios, drastic variation in the appearance of the same location depending on the time of the day, weather, season, and more importantly, the prediction is made from a single image possibly having only a few geo-locating cues. For these reasons, most existing works are restricted to specific cities, imagery, or worldwide landmarks. In this work, we focus on developing an efficient solution to planet-scale single-image geo-localization. To this end, we propose TransLocator, a unified dual-branch transformer network that attends to tiny details over the entire image and produces robust feature representation under extreme appearance variations. TransLocator takes an RGB image and its semantic segmentation map as inputs, interacts between its two parallel branches after each transformer layer, and simultaneously performs geo-localization and scene recognition in a multi-task fashion. We evaluate TransLocator on four benchmark datasets - Im2GPS, Im2GPS3k, YFCC4k, YFCC26k and obtain 5.5%, 14.1%, 4.9%, 9.9% continent-level accuracy improvement over the state-of-the-art. TransLocator is also validated on real-world test images and found to be more effective than previous methods."

"Although a plethora of architectural variants for deep classification has been introduced over time, recent works have found empirical evidence towards similarities in their training process. It has been hypothesized that neural networks converge not only to similar representations, but also exhibit a notion of empirical agreement on which data instances are learned first. Following in the latter works’ footsteps, we define a metric to quantify the relationship between such classification agreement over time, and posit that the agreement phenomenon can be mapped to core statistics of the investigated dataset. We empirically corroborate this hypothesis across the CIFAR10, Pascal, ImageNet and KTH-TIPS2 datasets. Our findings indicate that agreement seems to be independent of specific architectures, training hyper-parameters or labels, albeit follows an ordering according to image statistics."

"Recently, most handwritten mathematical expression recognition (HMER) methods adopt the encoder-decoder networks, which directly predict the markup sequences from formula images with the attention mechanism. However, such methods may fail to accurately read formulas with complicated structure or generate long markup sequences, as the attention results are often inaccurate due to the large variance of writing styles or spatial layouts. To alleviate this problem, we propose an unconventional network for HMER named Counting-Aware Network (CAN), which jointly optimizes two tasks: HMER and symbol counting. Specifically, we design a weakly-supervised counting module that can predict the number of each symbol class without the symbol-level position annotations, and then plug it into a typical attention-based encoder-decoder model for HMER. Experiments on the benchmark datasets for HMER validate that both joint optimization and counting results are beneficial for correcting the prediction errors of encoder-decoder models, and CAN consistently outperforms the state-of-the-art methods. In particular, compared with an encoder-decoder model for HMER, the extra time cost caused by the proposed counting module is marginal. The source code is available at https://github.com/LBH1024/CAN."