ICCV 2023 Papers

Deep neural networks are vulnerable to universal adversarial perturbation (UAP), an instance-agnostic perturbation capable of fooling the target model for most samples. Compared to instance-specific adversarial examples, UAP is more challenging as it needs to generalize across various samples and models. In this paper, we examine the serious dilemma of UAP generation methods from a generalization perspective -- the gradient vanishing problem using small-batch stochastic gradient optimization and the local optima problem using large-batch optimization. To address these problems, we propose a simple and effective method called Stochastic Gradient Aggregation (SGA), which alleviates the gradient vanishing and escapes from poor local optima at the same time. Specifically, SGA employs the small-batch training to perform multiple iterations of inner pre-search. Then, all the inner gradients are aggregated as a one-step gradient estimation to enhance the gradient stability and reduce quantization errors. Extensive experiments on the standard ImageNet dataset demonstrate that our method significantly enhances the generalization ability of UAP and outperforms other state-of-the-art methods. The code is available at https://github.com/liuxuannan/Stochastic-Gradient-Aggregation.

Backdoor defense, which aims to detect or mitigate the effect of malicious triggers introduced by attackers, is becoming increasingly critical for machine learning security and integrity. Fine-tuning based on benign data is a natural defense to erase the backdoor effect in a backdoored model. However, recent studies show that, given limited benign data, vanilla fine-tuning has poor defense performance. In this work, we firstly investigate the vanilla fine-tuning process for backdoor mitigation from the neuron weight perspective, and find that backdoor-related neurons are only slightly perturbed in the vanilla fine-tuning process, which explains its poor backdoor defense performance. To enhance the fine-tuning based defense, inspired by the observation that the backdoor-related neurons often have larger weight norms, we propose FT-SAM, a novel backdoor defense paradigm that aims to shrink the norms of backdoorrelated neurons by incorporating sharpness-aware minimization with fine-tuning. We demonstrate the effectiveness of our method on several benchmark datasets and network architectures, where it achieves state-of-the-art defense performance, and provide extensive analysis to reveal the FTSAM's mechanism. Overall, our work provides a promising avenue for improving the robustness of machine learning models against backdoor attacks. Codes are available at https://github.com/SCLBD/BackdoorBench.

Adversarial robustness is a research area that has recently received a lot of attention in the quest for trustworthy artificial intelligence. However, recent works on adversarial robustness have focused on supervised learning where it is assumed that labeled data is plentiful. In this paper, we investigate semi -supervised adversarial training where labeled data is scarce. We derive two upper bounds for the robust risk and propose a regularization term for unlabeled data motivated by these two upper bounds. Then, we develop a semi-supervised adversarial training algorithm that combines the proposed regularization term with knowledge distillation using a semi-supervised teacher. Our experiments show that our proposed algorithm achieves state-of-the-art performance with significant margins compared to existing algorithms. In particular, compared to supervised learning algorithms, performance of our proposed algorithm is not much worse even when the amount of labeled data is very small. For example, our algorithm with only 8% labeled data is comparable to supervised adversarial training algorithms that use all labeled data, both in terms of standard and robust accuracies on CIFAR-10.

As a mainstream framework in the field of semi-supervised learning (SSL), self-training via pseudo labeling and its variants have witnessed impressive progress in semi-supervised semantic segmentation with the recent advance of deep neural networks. However, modern self-training based SSL algorithms use a pre-defined constant threshold to select unlabeled pixel samples that contribute to the training, thus failing to be compatible with different learning difficulties of variant categories and different learning status of the model. To address these issues, we propose Enhanced Soft Label (ESL), a curriculum learning approach to fully leverage the high-value supervisory signals implicit in the untrustworthy pseudo label. ESL believes that pixels with unconfident predictions can be pretty sure about their belonging to a subset of dominant classes though being arduous to determine the exact one. It thus contains a Dynamic Soft Label (DSL) module to dynamically maintain the high probability classes, keeping the label "soft" so as to make full use of the high entropy prediction. However, the DSL itself will inevitably introduce ambiguity between dominant classes, thus blurring the classification boundary. Therefore, we further propose a pixel-to-part contrastive learning method cooperated with an unsupervised object part grouping mechanism to improve its ability to distinguish between different classes. Extensive experimental results on Pascal VOC 2012 and Cityscapes show that our approach achieves remarkable improvements over existing state-of-the-art approaches.

Deep Neural Networks (DNNs) have revolutionized visual classification tasks over the last decade. The training phase of deep-learning-based algorithms, however, often requires a vast amount of reliable annotated data. While reliability collecting such amount of labeled data usually yields to an exhaustive, expensive process, for many applications, acquiring massive datasets with imperfect annotations is straightforward. For instance, crawling search engines and online websites can generate a boatload amount of noisy labeled data. Hence, solving the problem of learning with noisy labels (LNL) is of paramount importance. Traditional LNL methods have successfully handled datasets with artificially injected noise, but they still fall short of adequately handling real-world noise. With the increasing use of meta-learning in the diverse fields of machine learning, researchers have tried to leverage auxiliary small clean datasets to meta-correct the training labels. Nonetheless, existing meta-label correction approaches are not fully exploiting their potential. In this study, we propose EMLC, an enhanced meta-label correction approach for the LNL problem. We re-examine the meta-learning process and introduce faster and more accurate meta-gradient derivations. We propose a novel teacher architecture tailored explicitly for the LNL problem, equipped with novel training objectives. EMLC outperforms prior approaches and achieves state-of-the-art results in all standard benchmarks. Notably, EMLC enhances the previous art on the noisy real-world dataset Clothing1M by 0.87%. Our publicly available code can be found at the following link: https://github.com/iccv23anonymous/Enhanced-Meta-Label-Correction

We propose an Unsupervised Domain Adaptation (UDA) method by making use of Energy-Based Learning (EBL) and demonstrate 1. EBL can be used to improve the instance selection for a self-training task on the unlabelled target domain, and 2. alignment and normalizing energy scores can learn domain-invariant representations. For the former, we show that an energy-based selection criterion can be used to model instance selections by mimicking the joint distribution between data and predictions in the target domain. As per learning domain invariant representations, we show that stable domain alignment can be achieved by a combined energy alignment and an energy normalization process. We implement our method in consistent with the vision-transformer (ViT) backbone and empirically show that our proposed method can outperform state-of-the-art ViT based UDA methods on diverse benchmarks (DomainNet, OfficeHome, and VISDA2017).

Erroneous feature matches have severe impact on subsequent camera pose estimation and often require additional, time-costly measures, like RANSAC, for outlier rejection. Our method tackles this challenge by addressing feature matching and pose optimization jointly. To this end, we propose a graph attention network to predict image correspondences along with confidence weights. The resulting matches serve as weighted constraints in a differentiable pose estimation. Training feature matching with gradients from pose optimization naturally learns to down-weight outliers and boosts pose estimation on image pairs compared to SuperGlue by 6.7% on ScanNet. At the same time, it reduces the pose estimation time by over 50% and renders RANSAC iterations unnecessary. Moreover, we integrate information from multiple views by spanning the graph across multiple frames to predict the matches all at once. Multi-view matching combined with end-to-end training improves the pose estimation metrics on Matterport3D by 18.5% compared to SuperGlue.

In this work, we present an end-to-end framework for camera-based 3D multi-object tracking, called DQTrack. To avoid heuristic design in detection-based trackers, recent query-based approaches deal with identity-agnostic detection and identity-aware tracking in a single embedding. However, it brings inferior performance because of the inherent representation conflict. To address this issue, we decouple the single embedding into separated queries, i.e., object query and track query. Unlike previous detection-based and query-based methods, the decoupled-query paradigm utilizes task-specific queries and still maintains the compact pipeline without complex post-processing. Moreover, the learnable association and temporal update are designed to provide differentiable trajectory association and frame-by-frame query update, respectively. The proposed DQTrack is demonstrated to achieve consistent gains in various benchmarks, outperforming all previous tracking-by-detection and learning-based methods on the nuScenes dataset.

Classifier guidance---using the gradients of an image classifier to steer the generations of a diffusion model---has the potential to dramatically expand the creative control over image generation and editing. However, currently classifier guidance requires either training new noise-aware models to obtain accurate gradients or using a one-step denoising approximation of the final generation, which leads to misaligned gradients and sub-optimal control. We highlight this approximation's shortcomings and propose a novel guidance method: Direct Optimization of Diffusion Latents (DOODL), which enables plug-and-play guidance by optimizing diffusion latents w.r.t. the gradients of a pre-trained classifier on the true generated pixels, using an invertible diffusion process to achieve memory-efficient backpropagation. Showcasing the potential of more precise guidance, DOODL outperforms one-step classifier guidance on computational and human evaluation metrics across different forms of guidance: using CLIP guidance to improve generations of complex prompts from DrawBench, using fine-grained visual classifiers to expand the vocabulary of Stable Diffusion, enabling image-conditioned generation with a CLIP visual encoder, and improving image aesthetics using an aesthetic scoring network.

We propose Encyclopedic-VQA, a large scale visual question answering (VQA) dataset featuring visual questions about detailed properties of fine-grained categories and instances. It contains 221k unique question+answer pairs each matched with (up to) 5 images, resulting in a total of 1M VQA samples. Moreover, our dataset comes with a controlled knowledge base derived from Wikipedia, marking the evidence to support each answer. Empirically, we show that our dataset poses a hard challenge for large vision+language models as they perform poorly on our dataset: PaLI [9] is state-of-the-art on OK-VQA [29], yet it only achieves 13.0% accuracy on our dataset. Moreover, we experimentally show that progress on answering our encyclopedic questions can be achieved by augmenting large models with a mechanism that retrieves relevant information for the knowledge base. An oracle experiment with perfect retrieval achieves 87.0% accuracy on the single-hop portion of our dataset, and an automatic retrieval- augmented prototype yields 48.8%. We believe that our dataset enables future research on retrieval-augmented vision+language models.

Deep neural networks have achieved remarkable progress in enhancing low-light images by improving their brightness and eliminating noise. However, most existing methods construct end-to-end mapping networks heuristically, neglecting the intrinsic prior of image enhancement task and lacking transparency and interpretability. Although some unfolding solutions have been proposed to relieve these issues, they rely on proximal operator networks that deliver ambiguous and implicit priors. In this work, we propose a paradigm for low-light image enhancement that explores the potential of customized learnable priors to improve the transparency of the deep unfolding paradigm.Motivated by the powerful feature representation capability of Masked Autoencoder (MAE), we customize MAE-based illumination and noise priors and redevelop them from two perspectives: 1) structure flow: we train the MAE from a normal-light image to its illumination properties and then embed it into the proximal operator design of the unfolding architecture; and 2) optimization flow: we train MAE from a normal-light image to its gradient representation and then employ it as a regularization term to constrain noise in the model output. These designs improve the interpretability and representation capability of the model. Extensive experiments on multiple low-light image enhancement datasets demonstrate the superiority of our proposed paradigm over state-of-the-art methods. Code is available at https://github.com/zheng980629/CUE.

Listener head generation centers on generating non-verbal behaviors (e.g., smile) of a listener in reference to the information delivered by a speaker. A significant challenge when generating such responses is the non-deterministic nature of fine-grained facial expressions during a conversation, which varies depending on the emotions and attitudes of both the speaker and the listener. To tackle this problem, we propose the Emotional Listener Portrait (ELP), which treats each fine-grained facial motion as a composition of several discrete motion-codewords and explicitly models the probability distribution of the motions under different emotional contexts in conversation. Benefiting from the "explicit" and "discrete" design, our ELP model can not only automatically generate natural and diverse responses toward a given speaker via sampling from the learned distribution but also generate controllable responses with a predetermined attitude. Under several quantitative metrics, our ELP exhibits significant improvements compared to previous methods.

Speech-driven 3D face animation aims to generate realistic facial expressions that match the speech content and emotion. However, existing methods often neglect emotional facial expressions or fail to disentangle them from speech content. To address this issue, this paper proposes an end-to-end neural network to disentangle different emotions in speech so as to generate rich 3D facial expressions. Specifically, we introduce the emotion disentangling encoder (EDE) to disentangle the emotion and content in the speech by cross-reconstructed speech signals with different emotion labels. Then an emotion-guided feature fusion decoder is employed to generate a 3D talking face with enhanced emotion. The decoder is driven by the disentangled identity, emotional, and content embeddings so as to generate controllable personal and emotional styles. Finally, considering the scarcity of the 3D emotional talking face data, we resort to the supervision of facial blendshapes, which enables the reconstruction of plausible 3D faces from 2D emotional data, and contribute a large-scale 3D emotional talking face dataset (3D-ETF) to train the network. Our experiments and user studies demonstrate that our approach outperforms state-of-the-art methods and exhibits more diverse facial movements. We recommend watching the supplementary video: https://ziqiaopeng. github.io/emotalk

Visual Emotion Analysis (VEA) aims at predicting people's emotional responses to visual stimuli. This is a promising, yet challenging, task in affective computing, which has drawn increasing attention in recent years. Most of the existing work in this area focuses on feature design, while little attention has been paid to dataset construction. In this work, we introduce EmoSet, the first large-scale visual emotion dataset annotated with rich attributes, which is superior to existing datasets in four aspects: scale, annotation richness, diversity, and data balance. EmoSet comprises 3.3 million images in total, with 118,102 of these images carefully labeled by human annotators, making it five times larger than the largest existing dataset. EmoSet includes images from social networks, as well as artistic images, and it is well balanced between different emotion categories. Motivated by psychological studies, in addition to emotion category, each image is also annotated with a set of describable emotion attributes: brightness, colorfulness, scene type, object class, facial expression, and human action, which can help understand visual emotions in a precise and interpretable way. The relevance of these emotion attributes is validated by analyzing the correlations between them and visual emotion, as well as by designing an attribute module to help visual emotion recognition. We believe EmoSet will bring some key insights and encourage further research in visual emotion analysis and understanding. Project page: https://vcc.tech/EmoSet.

Neural Architecture Search (NAS) has shown promising performance in the automatic design of vision transformers (ViT) exceeding 1G FLOPs. However, designing lightweight and low-latency ViT models for diverse mobile devices remains a big challenge. In this work, we propose ElasticViT, a two-stage NAS approach that trains a high-quality ViT supernet over a very large search space for covering a wide range of mobile devices, and then searches an optimal sub-network (subnet) for direct deployment. However, current supernet training methods that rely on uniform sampling suffer from the gradient conflict issue: the sampled subnets can have vastly different model sizes (e.g., 50M vs. 2G FLOPs), leading to different optimization directions and inferior performance. To address this challenge, we propose two novel sampling techniques: complexity-aware sampling and performance-aware sampling. Complexity-aware sampling limits the FLOPs difference among the subnets sampled across adjacent training steps, while covering different-sized subnets in the search space. Performance-aware sampling further selects subnets that have good accuracy, which can reduce gradient conflicts and improve supernet quality. Our discovered models, ElasticViT models, achieve top-1 accuracy from 67.2% to 80.0% on ImageNet from 60M to 800M FLOPs without extra retraining, outperforming all prior CNNs and ViTs in terms of accuracy and latency. Our tiny and small models are also the first ViT models that surpass state-of-the-art CNNs with significantly lower latency on mobile devices. For instance, ElasticViT-S1 runs 2.62x faster than EfficientNet-B0 with 0.1% higher accuracy.

Capturing high-dimensional social interactions and feasible futures is essential for predicting trajectories. To address this complex nature, several attempts have been devoted to reducing the dimensionality of the output variables via parametric curve fitting such as the Bezier curve and B-spline function. However, these functions, which originate in computer graphics fields, are not suitable to account for socially acceptable human dynamics. In this paper, we present EigenTrajectory (ET), a trajectory prediction approach that uses a novel trajectory descriptor to form a compact space, known here as ET space, in place of Euclidean space, for representing pedestrian movements. We first reduce the complexity of the trajectory descriptor via a low-rank approximation. We transform the pedestrians' history paths into our ET space represented by spatio-temporal principle components, and feed them into off-the-shelf trajectory forecasting models. The inputs and outputs of the models as well as social interactions are all gathered and aggregated in the corresponding ET space. Lastly, we propose a trajectory anchor-based refinement method to cover all possible futures in the proposed ET. Extensive experiments demonstrate that our EigenTrajectory predictor can significantly improve both the prediction accuracy and reliability of existing trajectory forecasting models on public benchmarks, indicating that the proposed descriptor is suited to represent pedestrian behaviors. Code is publicly available at https://github.com/inhwanbae/EigenTrajectory.

Visual Place Recognition is a task that aims to predict the place of an image (called query) based solely on its visual features. This is typically done through image retrieval, where the query is matched to the most similar images from a large database of geotagged photos, using learned global descriptors. A major challenge in this task is recognizing places seen from different viewpoints. To overcome this limitation, we propose a new method, called EigenPlaces, to train our neural network on images from different point of views, which embeds viewpoint robustness into the learned global descriptors. The underlying idea is to cluster the training data so as to explicitly present the model with different views of the same points of interest. The selection of this points of interest is done without the need for extra supervision. We then present experiments on the most comprehensive set of datasets in literature, finding that EigenPlaces is able to outperform previous state of the art on the majority of datasets, while requiring 60% less GPU memory for training and using 50% smaller descriptors. The code and trained models for EigenPlaces are available at https://github.com/gmberton/EigenPlaces, while results with any other baseline can be computed with the codebase at https://github.com/gmberton/auto_VPR.

Video-language pre-training (VLP) has become increasingly important due to its ability to generalize to various vision and language tasks. However, existing egocentric VLP frameworks utilize separate video and language encoders and learn task-specific cross-modal information only during fine-tuning, limiting the development of a unified system. In this work, we introduce the second generation of egocentric video-language pre-training (EgoVLPv2), a significant improvement from the previous generation, by incorporating cross-modal fusion directly into the video and language backbones. EgoVLPv2 learns strong video-text representation during pre-training and reuses the cross-modal attention modules to support different downstream tasks in a flexible and efficient manner, reducing fine-tuning costs. Moreover, our proposed fusion in the backbone strategy is more lightweight and compute-efficient than stacking additional fusion-specific layers. Extensive experiments on a wide range of VL tasks demonstrate the effectiveness of EgoVLPv2 by achieving consistent state-of-the-art performance over strong baselines across all downstream.

To enable progress towards egocentric agents capable of understanding everyday tasks specified in natural language, we propose a benchmark and a synthetic dataset called Egocentric Task Verification (EgoTV). The goal in EgoTV is to verify the execution of tasks from egocentric videos based on the natural language description of these tasks. EgoTV contains pairs of videos and their task descriptions for multi-step tasks -- these tasks contain multiple sub-task decompositions, state changes, object interactions, and sub-task ordering constraints. In addition, EgoTV also provides abstracted task descriptions that contain only partial details about ways to accomplish a task. Consequently, EgoTV requires causal, temporal, and compositional reasoning of video and language modalities, which is missing in existing datasets. We also find that existing vision-language models struggle at such all round reasoning needed for task verification in EgoTV. Inspired by the needs of EgoTV, we propose a novel Neuro-Symbolic Grounding (NSG) approach that leverages symbolic representations to capture the compositional and temporal structure of tasks. We demonstrate NSG's capability towards task tracking and verification on our EgoTV dataset and a real-world dataset derived from CrossTask (CTV). We open-source the EgoTV and CTV datasets and the NSG model for future research on egocentric assistive agents.