ICCV 2023 Papers

Bag-based multiple instance learning (MIL) methods have become the mainstream for Whole Slide Image (WSI) classification. However, there are still three important issues that have not been fully addressed: (1) positive bags with a low positive instance ratio are prone to the influence of a large number of negative instances; (2) the correlation between local and global features of pathology images has not been fully modeled; and (3) there is a lack of effective information interaction between different magnifications. In this paper, we propose MILBooster, a powerful dual-scale multi-stage MIL framework to address these issues from the perspectives of distribution, correlation, and magnification. Specifically, to address issue (1), we propose a plug-and-play bag filter that effectively increases the positive instance ratio of positive bags. For issue (2), we propose a novel window-based Transformer architecture called PiceBlock to model the correlation between local and global features of pathology images. For issue (3), we propose a dual-branch architecture to process different magnifications and design an information interaction module called Scale Mixer for efficient information interaction between them. We conducted extensive experiments on four clinical WSI classification tasks using three datasets. MILBooster achieved new state-of-the-art performance on all these tasks. Codes will be available.

Although deep learning has significantly facilitated the progress of single image super-resolution (SISR) in recent years, it still hits bottlenecks to further improve SR performance with the continuous growth of model scale. Therefore, one of the hotspots in the field is to construct efficient SISR models by elevating the effectiveness of feature representation. In this work, we present a straightforward and generic approach for feature enhancement that can effectively promote the performance of SR models, dubbed partial channel shifting (PCS). Specifically, it is inspired by the temporal shifting in video understanding and displaces part of the channels along the spatial dimensions, thus allowing the effective receptive field to be amplified and the feature diversity to be augmented at almost zero cost. Also, it can be assembled into off-the-shelf models as a plug-and-play component for performance boosting without extra network parameters and computational overhead. However, regulating the features with PCS encounters some issues, like shifting directions and amplitudes, proportions, and patterns of shifted channels, etc. We impose some technical constraints on the issues to simplify the general channel shifting. Extensive and throughout experiments illustrate that the PCS indeed enlarges the effective receptive field, augments the feature diversity for efficiently enhancing SR recovery, and can endow obvious performance gains to existing models.

Semantic segmentation is a computer vision task that associates a label with each pixel in an image. Modern approaches tend to introduce class embeddings into semantic segmentation for deeply utilizing category semantics, and regard supervised class masks as final predictions. In this paper, we explore the mechanism of class embeddings and have an insight that more explicit and meaningful class embeddings can be generated based on class masks purposely. Following this observation, we propose ECENet, a new segmentation paradigm, in which class embeddings are obtained and enhanced explicitly during interacting with multi-stage image features. Based on this, we revisit the traditional decoding process and explore inverted information flow between segmentation masks and class embeddings. Furthermore, to ensure the discriminability and informativity of features from backbone, we propose a Feature Reconstruction module, which combines intrinsic and diverse branches together to ensure the concurrence of diversity and redundancy in features. Experiments show that our ECENet outperforms its counterparts on the ADE20K dataset with much less computational cost and achieves new state-of-the-art results on PASCAL-Context dataset. The code will be released at https://gitee.com/mindspore/models and https://github.com/Carol-lyh/ECENet.

In this paper, we address the problem of weakly supervised Audio-Visual Video Parsing (AVVP), where the goal is to temporally localize events that are audible or visible and simultaneously classify them into known event categories. This is a challenging task, as we only have access to the video-level event labels during training but need to predict event labels at the segment level during evaluation. Existing multiple-instance learning (MIL) based methods use a form of attentive pooling over segment-level predictions. These methods only optimize for a subset of most discriminative segments that satisfy the weak-supervision constraints, which miss identifying positive segments. To address this, we focus on improving the proportion of positive segments detected in a video. To this end, we model the number of positive segments in a video as a latent variable and show that it can be modeled as Poisson binomial distribution over segment-level predictions, which can be computed exactly. Given the absence of fine-grained supervision, we propose an Expectation-Maximization approach to learn the model parameters by maximizing the evidence lower bound (ELBO). We iteratively estimate the minimum positive segments in a video and refine them to capture more positive segments. We conducted extensive experiments on AVVP tasks to evaluate the effectiveness of our proposed approach, and the results clearly demonstrate that it increases the number of positive segments captured compared to existing methods. Additionally, our experiments on Temporal Action Localization (TAL) demonstrate the potential of our method for generalization to similar MIL tasks.

While the field of multi-modal learning keeps growing fast, the deficiency of the standard joint training paradigm has become clear through recent studies. They attribute the sub-optimal performance of the jointly trained model to the modality competition phenomenon. Existing works attempt to improve the jointly trained model by modulating the training process. Despite their effectiveness, those methods can only apply to late fusion models. More importantly, the mechanism of the modality competition remains unexplored. In this paper, we first propose an adaptive gradient modulation method that can boost the performance of multi-modal models with various fusion strategies. Extensive experiments show that our method surpasses all existing modulation methods. Furthermore, to have a quantitative understanding of the modality competition and the mechanism behind the effectiveness of our modulation method, we introduce a novel metric to measure the competition strength. This metric is built on the mono-modal concept, a function that is designed to represent the competition-less state of a modality. Through systematic investigation, our results confirm the intuition that the modulation encourages the model to rely on the more informative modality. In addition, we find that the jointly trained model typically has a preferred modality on which the competition is weaker than other modalities. However, this preferred modality need not dominate others. Our code will be available at https://github.com/lihong2303/AGM_ICCV2023.

Real-world data tends to follow a long-tailed distribution, where the class imbalance results in dominance of the head classes during training. In this paper, we propose a frustratingly simple but effective step-wise learning framework to gradually enhance the capability of the model in detecting all categories of long-tailed datasets. Specifically, we build smooth-tail data where the long-tailed distribution of categories decays smoothly to correct the bias towards head classes. We pre-train a model on the whole long-tailed data to preserve discriminability between all categories. We then fine-tune the class-agnostic modules of the pre-trained model on the head class dominant replay data to get a head class expert model with improved decision boundaries from all categories. Finally, we train a unified model on the tail class dominant replay data while transferring knowledge from the head class expert model to ensure accurate detection of all categories. Extensive experiments on long-tailed datasets LVIS v0.5 and LVIS v1.0 demonstrate the superior performance of our method, where we can improve the AP with ResNet-50 backbone from 27.0% to 30.3% AP, and especially for the rare categories from 15.5% to 24.9% AP. Our best model using ResNet-101 backbone can achieve 30.7% AP, which suppresses all existing detectors using the same backbone.

Class prototype construction and matching are core aspects of few-shot action recognition. Previous methods mainly focus on designing spatiotemporal relation modeling modules or complex temporal alignment algorithms. Despite the promising results, they ignored the value of class prototype construction and matching, leading to unsatisfactory performance in recognizing similar categories in every task. In this paper, we propose GgHM, a new framework with Graph-guided Hybrid Matching. Concretely, we learn task-oriented features by the guidance of a graph neural network during class prototype construction, optimizing the intra- and inter-class feature correlation explicitly. Next, we design a hybrid matching strategy, combining frame-level and tuple-level matching to classify videos with multivariate styles. We additionally propose a learnable dense temporal modeling module to enhance the video feature temporal representation to build a more solid foundation for the matching process. GgHM shows consistent improvements over other challenging baselines on several few-shot datasets, demonstrating the effectiveness of our method. The code will be publicly available at https://github.com/jiazheng-xing/GgHM.

Plentiful adversarial attack researches have revealed the fragility of deep neural networks (DNNs), where the imperceptible perturbations can cause drastic changes in the output. Among the diverse types of attack methods, gradient-based attacks are powerful and easy to implement, arousing wide concern for the security problem of DNNs. However, under the black-box setting, the existing gradient-based attacks have much trouble in breaking through DNN models with defense technologies, especially those adversarially trained models. To make adversarial examples more transferable, in this paper, we explore the fluctuation phenomenon on the plus-minus sign of the adversarial perturbations' pixels during the generation of adversarial examples, and propose an ingenious Gradient Relevance Attack (GRA). Specifically, two gradient relevance frameworks are presented to better utilize the information in the neighborhood of the input, which can correct the update direction adaptively. Then we adjust the update step at each iteration with a decay indicator to counter the fluctuation. Experiment results on a subset of the ILSVRC 2012 validation set forcefully verify the effectiveness of GRA. Furthermore, the attack success rates of 68.7% and 64.8% on Tencent Cloud and Baidu AI Cloud further indicate that GRA can craft adversarial examples with the ability to transfer across both datasets and model architectures. Code is released at https://github.com/RYC-98/GRA.

Image retrieval-based cross-view localization methods often lead to very coarse camera pose estimation, due to the limited sampling density of the database satellite images. In this paper, we propose a method to increase the accuracy of a ground camera's location and orientation by estimating the relative rotation and translation between the ground-level image and its matched/retrieved satellite image. Our approach designs a geometry-guided cross-view transformer that combines the benefits of conventional geometry and learnable cross-view transformers to map the ground-view observations to an overhead view. Given the synthesized overhead view and observed satellite feature maps, we construct a neural pose optimizer with strong global information embedding ability to estimate the relative rotation between them. After aligning their rotations, we develop an uncertainty-guided spatial correlation to generate a probability map of the vehicle locations, from which the relative translation can be determined. Experimental results demonstrate that our method significantly outperforms the state-of-the-art. Notably, the likelihood of restricting the vehicle lateral pose to be within 1m of its Ground Truth (GT) value on the cross-view KITTI dataset has been improved from 35.54% to 76.44%, and the likelihood of restricting the vehicle orientation to be within 1 degree of its GT value has been improved from 19.64% to 99.10%.

Personalized federated learning (PFL) reduces the impact of non-independent and identically distributed (non-IID) data among clients by allowing each client to train a personalized model when collaborating with others. A key question in PFL is to decide which parameters of a client should be localized or shared with others. In current mainstream approaches, all layers that are sensitive to non-IID data (such as classifier layers) are generally personalized. The reasoning behind this approach is understandable, as localizing parameters that are easily influenced by non-IID data can prevent potential negative effects of collaboration. However, we believe that this approach is too conservative for collaboration. For example, for a certain client, even if its parameters are easily influenced by non-IID data, it can still benefit by sharing these parameters with clients having similar data distribution. This observation emphasizes the importance of considering not only the sensitivity to non-IID data but also the similarity of data distribution when determining which parameters should be localized in PFL. This paper introduces a novel guideline for client collaboration in PFL. Unlike existing approaches that prohibit all collaboration of sensitive parameters, our guideline allows clients to share more parameters with others, leading to improved model performance. Additionally, we propose a new PFL method named FedCAC, which employs a quantitative metric to evaluate each parameter's sensitivity to non-IID data and carefully selects collaborators based on this evaluation. Experimental results demonstrate that FedCAC enables clients to share more parameters with others, resulting in superior performance compared to state-of-the-art methods, particularly in scenarios where clients have diverse distributions.

While 3D body reconstruction methods have made remarkable progress recently, it remains difficult to acquire the sufficiently accurate and numerous 3D supervisions required for training. In this paper, we propose KNOWN, a framework that effectively utilizes body KNOWledge and uNcertainty modeling to compensate for insufficient 3D supervisions. KNOWN exploits a comprehensive set of generic body constraints derived from well-established body knowledge. These generic constraints precisely and explicitly characterize the reconstruction plausibility and enable 3D reconstruction models to be trained without any 3D data. Moreover, existing methods typically use images from multiple datasets during training, which can result in data noise (e.g., inconsistent joint annotation) and data imbalance (e.g., minority images representing unusual poses or captured from challenging camera views). KNOWN solves these problems through a novel probabilistic framework that models both aleatoric and epistemic uncertainty. Aleatoric uncertainty is encoded in a robust Negative Log-Likelihood (NLL) training loss, while epistemic uncertainty is used to guide model refinement. Experiments demonstrate that KNOWN's body reconstruction outperforms prior weakly-supervised approaches, particularly on the challenging minority images.

Deep learning methods have shown outstanding classification accuracy in medical imaging problems, which is largely attributed to the availability of large-scale datasets manually annotated with clean labels. However, given the high cost of such manual annotation, new medical imaging classification problems may need to rely on machine-generated noisy labels extracted from radiology reports. Indeed, many Chest X-Ray (CXR) classifiers have been modelled from datasets with noisy labels, but their training procedure is in general not robust to noisy-label samples, leading to sub-optimal models. Furthermore, CXR datasets are mostly multi-label, so current multi-class noisy-label learning methods cannot be easily adapted. In this paper, we propose a new method designed for noisy multi-label CXR learning, which detects and smoothly re-labels noisy samples from the dataset to be used in the training of common multi-label classifiers. The proposed method optimises a bag of multi-label descriptors (BoMD) to promote their similarity with the semantic descriptors produced by language models from multi-label image annotations. Our experiments on noisy multi-label training sets and clean testing sets show that our model has state-of-the-art accuracy and robustness in many CXR multi-label classification benchmarks, including a new benchmark that we propose to systematically assess noisy multi-label methods.

In MRI, images of the same contrast (e.g., T1) from the same subject can exhibit noticeable differences when acquired using different hardware, sequences, or scan parameters. These differences in images create a domain gap that needs to be bridged by a step called image harmonization, to process the images successfully using conventional or deep learning-based image analysis (e.g., segmentation). Several methods, including deep learning-based approaches, have been proposed to achieve image harmonization. However, they often require datasets from multiple domains for deep learning training and may still be unsuccessful when applied to images from unseen domains. To address this limitation, we propose a novel concept called 'Blind Harmonization', which utilizes only target domain data for training but still has the capability to harmonize images from unseen domains. For the implementation of blind harmonization, we developed BlindHarmony using an unconditional flow model trained on target domain data. The harmonized image is optimized to have a correlation with the input source domain image while ensuring that the latent vector of the flow model is close to the center of the Gaussian distribution. BlindHarmony was evaluated on both simulated and real datasets and compared to conventional methods. BlindHarmony demonstrated noticeable performance on both datasets, highlighting its potential for future use in clinical settings. The source code is available at: https://github.com/SNU-LIST/BlindHarmony

Text-driven localized editing of 3D objects is particularly difficult as locally mixing the original 3D object with the intended new object and style effects without distorting the object's form is not a straightforward process. To address this issue, we propose a novel NeRF-based model, Blending-NeRF, which consists of two NeRF networks: pretrained NeRF and editable NeRF. Additionally, we introduce new blending operations that allow Blending-NeRF to properly edit target regions which are localized by text. By using a pretrained vision-language aligned model, CLIP, we guide Blending-NeRF to add new objects with varying colors and densities, modify textures, and remove parts of the original object. Our extensive experiments demonstrate that Blending-NeRF produces naturally and locally edited 3D objects from various text prompts.

The great advancements of generative adversarial networks and face recognition models in computer vision have made it possible to swap identities on images from single sources. Although a lot of studies seems to have proposed almost satisfactory solutions, we notice previous methods still suffer from an identity-attribute entanglement that causes undesired attributes swapping because widely used identity encoders, e.g., ArcFace, have some crucial attribute biases owing to their pretraining on face recognition tasks. To address this issue, we design BlendFace, a novel identity encoder for face-swapping. The key idea behind BlendFace is training face recognition models on blended images whose attributes are replaced with those of another mitigates inter-personal biases such as hairsyles and head shapes. BlendFace feeds disentangled identity features into generators and guides generators properly as an identity loss function. Extensive experiments demonstrate that BlendFace improves the identity-attribute disentanglement in face-swapping models, maintaining a comparable quantitative performance to previous methods.

Black-box unsupervised domain adaptation (UDA) learns with source predictions of target data without accessing either source data or source models during training, and it has clear superiority in data privacy and flexibility in target network selection. However, the source predictions of target data are often noisy and training with them is prone to learning collapses. We propose BiMem, a bi-directional memorization mechanism that learns to remember useful and representative information to correct noisy pseudo labels on the fly, leading to robust black-box UDA that can generalize across different visual recognition tasks. BiMem constructs three types of memory, including sensory memory, short-term memory, and long-term memory, which interact in a bi-directional manner for comprehensive and robust memorization of learnt features. It includes a forward memorization flow that identifies and stores useful features and a backward calibration flow that rectifies features' pseudo labels progressively. Extensive experiments show that BiMem achieves superior domain adaptation performance consistently across various visual recognition tasks such as image classification, semantic segmentation and object detection.

Vision-Language (V-L) models trained with contrastive learning to align the visual and language modalities have been shown to be strong few-shot learners. Soft prompt learning is the method of choice for few-shot downstream adaption aiming to bridge the modality gap caused by the distribution shift induced by the new domain. While parameter-efficient, prompt learning still requires access to the model weights and can be computationally infeasible for large models with billions of parameters. To address these shortcomings, in this work, we describe a black-box method for V-L few-shot adaptation that (a) operates on pre-computed image and text features and hence works without access to the model's weights, (b) it is orders of magnitude faster at training time, (c) it is amenable to both supervised and unsupervised training, and (d) it can be even used to align image and text features computed from uni-modal models. To achieve this, we propose Linear Feature Alignment (LFA), a simple linear approach for V-L re-alignment in the target domain. LFA is initialized from a closed-form solution to a least-squares problem and then it is iteratively updated by minimizing a re-ranking loss. Despite its simplicity, our approach can even surpass soft-prompt learning methods as shown by extensive experiments on 11 image and 2 video datasets.

Vision-language navigation (VLN), which entails an agent to navigate 3D environments following human instructions, has shown great advances. However, current agents are built upon panoramic observations, which hinders their ability to perceive 3D scene geometry and easily leads to ambiguous selection of panoramic view. To address these limitations, we present a BEV Scene Graph (BSG), which leverages multi-step BEV representations to encode scene layouts and geometric cues of indoor environment under the supervision of 3D detection. During navigation, BSG builds a local BEV representation at each step and maintains a BEV-based global scene map, which stores and organizes all the online collected local BEV representations according to their topological relations. Based on BSG, the agent predicts a local BEV grid-level decision score and a global graph-level decision score, combined with a subview selection score on panoramic views, for more accurate action prediction. Our approach significantly outperforms state-of-the-art methods on REVERIE, R2R, and R4R, showing the potential of BEV perception in VLN.

Video-based human pose transfer is a video-to-video generation task that animates a plain source human image based on a series of target human poses. Considering the difficulties in transferring highly structural patterns on the garments and discontinuous poses, existing methods often generate unsatisfactory results such as distorted textures and flickering artifacts. To address these issues, we propose a novel Deformable Motion Modulation (DMM) that utilizes geometric kernel offset with adaptive weight modulation to simultaneously perform feature alignment and style transfer. Different from normal style modulation used in style transfer, the proposed modulation mechanism adaptively reconstructs smoothed frames from style codes according to the object shape through an irregular receptive field of view. To enhance the spatio-temporal consistency, we leverage bidirectional propagation to extract the hidden motion information from a warped image sequence generated by noisy poses. The proposed feature propagation significantly enhances the motion prediction ability by forward and backward propagation. Both quantitative and qualitative experimental results demonstrate superiority over the state-of-the-arts in terms of image fidelity and visual continuity. The source code is publicly available at github.com/rocketappslab/bdmm.