Search

Noninvasively and accurately predicting the epidermal growth factor receptor (EGFR) mutation status is a clinically vital problem. Moreover, further identifying the most suspicious area related to the EGFR mutation status can guide the biopsy to avoid false negatives. Deep learning methods based on computed tomography (CT) images may improve the noninvasive prediction of EGFR mutation status and potentially help clinicians guide biopsies by visual methods. Inspired by the potential inherent links between EGFR mutation status and invasiveness information, we hypothesized that the predictive performance of a deep learning network can be improved through extra utilization of the invasiveness information. Here, we created a novel explainable transformer network for EGFR classification named gated multiple instance learning transformer (GMILT) by integrating multi-instance learning and discriminative weakly supervised feature learning. Pathological invasiveness information was first introduced into the multitask model as embeddings. GMILT was trained and validated on a total of 512 patients with adenocarcinoma and tested on three datasets (the internal test dataset, the external test dataset, and The Cancer Imaging Archive (TCIA) public dataset). The performance (area under the curve (AUC) $=0.772$ on the internal test dataset) of GMILT exceeded that of previously published methods and radiomics-based methods (i.e., random forest and support vector machine) and attained a preferable generalization ability (AUC $=0.856$ in the TCIA test dataset and AUC $=0.756$ in the external dataset). A diameter-based subgroup analysis further verified the efficiency of our model (most of the AUCs exceeded 0.772) to noninvasively predict EGFR mutation status from computed tomography (CT) images. In addition, because our method also identified the “core area” of the most suspicious area related to the EGFR mutation status, it has the potential ability to guide biopsies.

Face identity editing (FIE) shows great value in AI content creation. Low-resolution FIE approaches have achieved tremendous progress, but high-quality FIE struggles. Two major challenges hinder higher-resolution and higher-performance development of FIE: lack of high-resolution dataset and unacceptable complexity forbidding for mobile platforms. To address both issues, we establish a novel large-scale, high-quality dataset tailored for FIE. Based on our SimSwap (Chen et al. 2020), we propose an upgraded version named SimSwap++ with significantly boosted model efficiency. SimSwap++ features two major innovations for high-performance model compression. First, a novel computational primitive named Conditional Dynamic Convolution (CD-Conv) is proposed to address the inefficiency of conditional schemes (e.g., AdaIN) in tiny models. CD-Conv achieves anisotropic processing and injection with significantly lower complexity compared to standard conditional operators, e.g., modulated convolution. Second, a Morphable Knowledge Distillation (MKD) is presented to further trim the overall model. Unlike conventional homogeneous teacher-student structures, MKD is designed to be heterogeneous and mutually compensable, endowing the student with the multi-path morphable property; thus, our student maximally inherits the teacher's knowledge after distillation while further reducing its complexity through structure re-parameterization. Extensive experiments demonstrate that our SimSwap++ achieves state-of-the-art performance (97.55% ID accuracy on FaceForensics++) with extremely low complexity (2.5 GFLOPs).

Multi-Source Domain Adaptation (MSDA) focuses on transferring the knowledge from multiple source domains to the target domain, which is a more practical and challenging problem compared to the conventional single-source domain adaptation. In this problem, it is essential to model multiple source domains and target domain jointly, and an effective domain combination scheme is also highly required. The graphical structure among different domains is useful to tackle these challenges, in which the interdependency among various instances/categories can be effectively modeled. In this work, we propose two types of graphical models, i.e. Conditional Random Field for MSDA (CRF-MSDA) and Markov Random Field for MSDA (MRF-MSDA), for cross-domain joint modeling and learnable domain combination. In a nutshell, given an observation set composed of a query sample and the semantic prototypes (i.e. representative category embeddings) on various domains, the CRF-MSDA model seeks to learn the joint distribution of labels conditioned on the observations. We attain this goal by constructing a relational graph over all observations and conducting local message passing on it. By comparison, MRF-MSDA aims to model the joint distribution of observations over different Markov networks via an energy-based formulation, and it can naturally perform label prediction by summing the joint likelihoods over several specific networks. Compared to the CRF-MSDA counterpart, the MRF-MSDA model is more expressive and possesses lower computational cost. We evaluate these two models on four standard benchmark data sets of MSDA with distinct domain shift and data complexity, and both models achieve superior performance over existing methods on all benchmarks. In addition, the analytical studies illustrate the effect of different model components and provide insights about how the cross-domain joint modeling performs.

Various methods have been proposed to defend against adversarial attacks. However, there is a lack of enough theoretical guarantee of the performance, thus leading to two problems: First, deficiency of necessary adversarial training samples might attenuate the normal gradient's back-propagation, which leads to overfitting and gradient masking potentially. Second, point-wise adversarial sampling offers an insufficient support region for adversarial data and thus cannot form a robust decision-boundary. To solve these issues, we provide a theoretical analysis to reveal the relationship between robust accuracy and the complexity of the training set in adversarial training. As a result, we propose a novel training scheme called Variational Adversarial Defense. Based on the distribution of adversarial samples, this novel construction upgrades the defend scheme from local point-wise to distribution-wise, yielding an enlarged support region for safeguarding robust training, thus possessing a higher promising to defense attacks. The proposed method features the following advantages: 1) Instead of seeking adversarial examples point-by-point (in a sequential way), we draw diverse adversarial examples from the inferred distribution; and 2) Augmenting the training set by a larger support region consolidates the smoothness of the decision boundary. Finally, the proposed method is analyzed via the Taylor expansion technique, which casts our solution with natural interpretability.

Automatic rib labeling and anatomical centerline extraction are common prerequisites for various clinical applications. Prior studies either use in-house datasets that are inaccessible to communities, or focus on rib segmentation that neglects the clinical significance of rib labeling. To address these issues, we extend our prior dataset (RibSeg) on the binary rib segmentation task to a comprehensive benchmark, named RibSeg v2, with 660 CT scans (15,466 individual ribs in total) and annotations manually inspected by experts for rib labeling and anatomical centerline extraction. Based on the RibSeg v2, we develop a pipeline including deep learning-based methods for rib labeling, and a skeletonization-based method for centerline extraction. To improve computational efficiency, we propose a sparse point cloud representation of CT scans and compare it with standard dense voxel grids. Moreover, we design and analyze evaluation metrics to address the key challenges of each task. Our dataset, code, and model are available online to facilitate open research at https://github.com/M3DV/RibSeg.

Supervised learning is a mainstay for large discriminative models in 3D computer vision, while large amounts of human-annotated data are the key to achieve state-of-the-art performance. This limitation is particularly notable for large-scale point cloud sequence segmentation tasks, because point-level annotations are very time-consuming and especially expensive. To overcome this challenge, we develop a novel semi-supervised framework for point cloud sequences segmentation. Specifically, we develop two kinds of pseudo labeling methods with extracting global semantic information from labeled frames and dynamic information from each sequence respectively. Then the two kinds of generated labels are combined as more robust pseudo labels (GD-Pseudo labels) for unlabeled frames. We finally apply an efficient iterative learning scheme to train a model with a small quantity of human-annotated data and large-scale pseudo-labeled data. Equipped with our framework, the model achieves significant performance improvement (+12—25 mIoU) on SemanticKITTI and Synthia when compared with frameworks that do not utilize large amounts of unlabeled data. Moreover, our method achieves comparable performance with only 20% annotated frames on SemanticKITTI to state-of-the-art models trained with 100% human-annotated frames.

Style-based GANs achieve state-of-the-art results for generating high-quality images, but lack explicit and precise control over camera poses. Recently proposed NeRF-based GANs have made great progress towards 3D-aware image generation. However, the methods either rely on convolution operators which are not rotationally invariant, or utilize complex yet suboptimal training procedures to integrate both NeRF and CNN sub-structures, yielding un-robust, low-quality images with a large computational burden. This article presents an upgraded version called CIPS-3D++, aiming at high-robust, high-resolution and high-efficiency 3D-aware GANs. On the one hand, our basic model CIPS-3D, encapsulated in a style-based architecture, features a shallow NeRF-based 3D shape encoder as well as a deep MLP-based 2D image decoder, achieving robust image generation/editing with rotation-invariance. On the other hand, our proposed CIPS-3D++, inheriting the rotational invariance of CIPS-3D, together with geometric regularization and upsampling operations, encourages high-resolution high-quality image generation/editing with great computational efficiency. Trained on raw single-view images, without any bells and whistles, CIPS-3D++ sets new records for 3D-aware image synthesis, with an impressive FID of 3.2 on FFHQ at the $1024\times 1024$1024×1024 resolution. In the meantime, CIPS-3D++ runs efficiently and enjoys a low GPU memory footprint so that it can be trained end-to-end on high-resolution images directly, in contrast to previous alternate/progressive methods. Based on the infrastructure of CIPS-3D++, we propose a 3D-aware GAN inversion algorithm named FlipInversion, which can reconstruct the 3D object from a single-view image. We also provide a 3D-aware stylization method for real images based on CIPS-3D++ and FlipInversion. In addition, we analyze the problem of mirror symmetry suffered in training, and solve it by introducing an auxiliary discriminator for the NeRF network. Overall, CIPS-3D++ provides a strong base model that can serve as a testbed for transferring GAN-based image editing methods from 2D to 3D.

To address the limitations of existing pruning methods in practical applications, such as the necessity of training from scratch with sparsity regularization or complex data-driven optimization, we set out from a novel perspective to explore parameter redundancy and accelerate deep CNNs. Precisely, we argue that channels revealing similar feature information have functional overlap and that each such similarity group can be reduced to a few representatives with little impact on the representational power of the model. After deriving an effective metric for evaluating channel similarity via probabilistic modeling, we introduce a similarity-based pruning framework based on hierarchical clustering.In particular, the proposed algorithm can be directly applied to all kinds of pre-trained CNN models for better trade-offs between latency and accuracy. Moreover, rather than relying on a pre-defined target structure, it automatically discovers resource-efficient ones out of the original model under given budgets, which is in the same flavor as NAS. Extensive experiments on benchmark datasets well demonstrate the superior performance of our approach over prior arts. On ImageNet, our pruned ResNet-50 with 30% FLOPs reduced outperforms the original model. We further extend our algorithm to a GAN-based generative model and achieve $2\times $ acceleration, showing its remarkable generalization capability and flexibility.

The key point for an experienced craftsman to repair broken objects effectively is that he must know about them deeply. Similarly, we believe that a model can capture rich geometry information from a shape/scene and generate discriminative representations if it is able to find distorted parts of shapes/scenes and restore them. Inspired by this observation, we propose a novel self-supervised 3D learning paradigm named learning by restoring broken shapes/scenes (collectively called 3D geometry). We first develop a destroy-method cluster, from which we sample methods to break some local parts of an object. Then the destroyed object and the normal object are both sent into a point cloud network to get representations, which are employed to segment points that belong to distorted parts and further reconstruct/restore them to normal. To perform better in these two associated pretext tasks, the model is constrained to capture useful object features, such as rich geometric and contextual information. The object representations learned by this self-supervised paradigm transfer well to different datasets and perform well on downstream classification, segmentation and detection tasks. Experimental results on shape datasets and scene datasets demonstrate that our method achieves state-of-the-art performance among unsupervised methods. We also show experimentally that pre-training with our framework significantly boosts the performance of supervised models.

We investigate GAN inversion problems of using pre-trained GANs to reconstruct real images. Recent methods for such problems typically employ a VGG perceptual loss to measure the difference between images. While the perceptual loss has achieved remarkable success in various computer vision tasks, it may cause unpleasant artifacts and is sensitive to changes in input scale. This paper delivers an important message that algorithm details are crucial for achieving satisfying performance. In particular, we propose two important but undervalued design principles: (i) not down-sampling the input of the perceptual loss to avoid high-frequency artifacts; and (ii) calculating the perceptual loss using convolutional features which are robust to scale. Integrating these designs derives the proposed framework, HRInversion, that achieves superior performance in reconstructing image details. We validate the effectiveness of HRInversion on a cross-domain image synthesis task and propose a post-processing approach named local style optimization (LSO) to synthesize clean and controllable stylized images. For the evaluation of the cross-domain images, we introduce a metric named ID retrieval which captures the similarity of face identities of stylized images to content images. We also test HRInversion on non-square images. Equipped with implicit neural representation, HRInversion applies to ultra-high resolution images with more than 10 million pixels. Furthermore, we show applications of style transfer and 3D-aware GAN inversion, paving the way for extending the application range of HRInversion.

Capturing the interactions of human articulations lies in the center of skeleton-based action recognition. Recent graph-based methods are inherently limited in the weak spatial context modeling capability due to fixed interaction pattern and inflexible shared weights of GCN. To address above problems, we propose the multi-view interactional graph network (MV-IGNet) which can construct, learn and infer multi-level spatial skeleton context, including view-level (global), group-level, joint-level (local) context, in a unified way. MV-IGNet leverages different skeleton topologies as multi-views to cooperatively generate complementary action features. For each view, separable parametric graph convolution (SPG-Conv) enables multiple parameterized graphs to enrich local interaction patterns, which provides strong graph-adaption ability to handle irregular skeleton topologies. We also partition the skeleton into several groups and then the higher-level group contexts including inter-group and intra-group, are hierarchically captured by above SPG-Conv layers. A simple yet effective global context adaption (GCA) module facilitates representative feature extraction by learning the input-dependent skeleton topologies. Compared to the mainstream works, MV-IGNet can be readily implemented while with smaller model size and faster inference. Experimental results show the proposed MV-IGNet achieves impressive performance on large-scale benchmarks: NTU-RGB+D and NTU-RGB+D 120.

Learning to re-identify or retrieve a group of people across non-overlapped camera systems has important applications in video surveillance. However, most existing methods focus on (single) person re-identification (re-id), ignoring the fact that people often walk in groups in real scenarios. In this work, we take a step further and consider employing context information for identifying groups of people, i.e., group re-id. On the one hand, group re-id is more challenging than single person re-id, since it requires both a robust modeling of local individual person appearance (with different illumination conditions, pose/viewpoint variations, and occlusions), as well as full awareness of global group structures (with group layout and group member variations). On the other hand, we believe that person re-id can be greatly enhanced by incorporating additional visual context from neighboring group members, a task which we formulate as group-aware (single) person re-id. In this paper, we propose a novel unified framework based on graph neural networks to simultaneously address the above two group-based re-id tasks, i.e., group re-id and group-aware person re-id. Specifically, we construct a context graph with group members as its nodes to exploit dependencies among different people. A multi-level attention mechanism is developed to formulate both intra-group and inter-group context, with an additional self-attention module for robust graph-level representations by attentively aggregating node-level features. The proposed model can be directly generalized to tackle group-aware person re-id using node-level representations. Meanwhile, to facilitate the deployment of deep learning models on these tasks, we build a new group re-id dataset which contains more than $3.8K$3.8K images with $1.5K$1.5K annotated groups, an order of magnitude larger than existing group re-id datasets. Extensive experiments on the novel dataset as well as three existing datasets clearly demonstrate the effectiveness of the proposed framework for both group-based re-id tasks.

We consider a new problem of adapting a human mesh reconstruction model to out-of-domain streaming videos, where the performance of existing SMPL-based models is significantly affected by the distribution shift represented by different camera parameters, bone lengths, backgrounds, and occlusions. We tackle this problem through online adaptation, gradually correcting the model bias during testing. There are two main challenges: First, the lack of 3D annotations increases the training difficulty and results in 3D ambiguities. Second, non-stationary data distribution makes it difficult to strike a balance between fitting regular frames and hard samples with severe occlusions or dramatic changes. To this end, we propose the Dynamic Bilevel Online Adaptation algorithm (DynaBOA). It first introduces the temporal constraints to compensate for the unavailable 3D annotations and leverages a bilevel optimization procedure to address the conflicts between multi-objectives. DynaBOA provides additional 3D guidance by co-training with similar source examples retrieved efficiently despite the distribution shift. Furthermore, it can adaptively adjust the number of optimization steps on individual frames to fully fit hard samples and avoid overfitting regular frames. DynaBOA achieves state-of-the-art results on three out-of-domain human mesh reconstruction benchmarks.

A fundamental bottleneck for achieving highly discriminative action representation is that local motion/appearance features are usually not semantic aligned. Namely, a local feature, such as a motion vector or motion trajectory, does not possess any attribute that indicates which moving body part or operated object it is associated with. This mostly leads to global feature pooling/representation learning methods that are often too coarse. Inspired by the recent success of end-to-end (pixel-to-pixel) deep convolutional neural networks (DCNNs), in this paper, we first propose a DCNN architecture, which maps a human centric image region onto human body part response maps. Based on these response maps, we propose a second DCNN, which achieves semantic-aligned feature representation learning. Prior knowledge that only a few parts are responsible for a certain action is also utilized by introducing a group (part) sparseness prior during feature learning. The learned semantic-aligned feature not only boosts the discriminative capability of action representation, but also possesses the good nature of robustness to pose variations and occlusions. Finally, an iterative mining method is employed for learning discriminative action primitive detectors. Extensive experiments on action recognition benchmarks demonstrate a superior recognition performance of the proposed framework. [ABSTRACT FROM AUTHOR]

This article investigates a low-rank representation--based graph, which can used in graph-based vision tasks including image segmentation and image annotation. It naturally fuses multiple types of image features in a framework named multitask low-rank affinity pursuit. Given the image patches described with multiple types of features, we aim at inferring a unified affinity matrix that implicitly encodes the relations among these patches. This is achieved by seeking the sparsity-consistent low-rank affinities from the joint decompositions of multiple feature matrices into pairs of sparse and low-rank matrices, the latter of which is expressed as the production of the image feature matrix and its corresponding image affinity matrix. The inference process is formulated as a minimization problem and solved efficiently with the augmented Lagrange multiplier method. Considering image patches as vertices, a graph can be built based on the resulted affinity matrix. Compared to previous methods, which are usually based on a single type of feature, the proposed method seamlessly integrates multiple types of features to jointly produce the affinity matrix in a single inference step. The proposed method is applied to graph-based image segmentation and graph-based image annotation. Experiments on benchmark datasets well validate the superiority of using multiple features over single feature and also the superiority of our method over conventional methods for feature fusion.

In this article, beyond solo-activity analysis for single object, we study the more complicated pair-activity recognition problem by exploring the relationship between two active objects based on their trajectory clues obtained from video sensor. Our contributions are three-fold. First, we design two sets of features for representing the pair-activities encoded as length-variable trajectory pairs. One set characterizes the strength of causality between two trajectories, for example, the causality ratio and feedback ratio based on the Granger Causality Test (GCT), and another set describes the style of causality between two trajectories, for example, the sampled frequency responses of the digital filter with these two trajectories as the input and output discrete signals respectively. These features along with conventional velocity and position features of a trajectory-pair are essentially of multi-modalities, and may be greatly different in scales and importance. To make full use of them, we then develop a novel feature fusing procedure to learn the coefficients for weighting these features by maximizing the discriminating power measured by weighted correlation. Finally, we collected a pair-activity database of five popular categories, each of which consists of about 170 instances. The extensive experiments on this database validate the effectiveness of the designed features for pair-activity representation, and also demonstrate that the proposed feature fusing procedure significantly boosts the pair-activity classification accuracy.

Scene parsing, fully labeling an image with each region corresponding to a label, is one of the core problems of computer vision. Previous methods to this problem usually rely on patch-level models trained from well labeled data. In this paper, we propose a weakly-supervised scene parsing algorithm that semantically parses a collection of images with multi-label, which is guided by the top-down category models and bottom-up local patch contexts across images that closely related segments usually have similar labels. Images are segmented to patches on multi-level and the contextual relations of patches are discovered via sparse representation by ℓ 1 minimization, based on which a graph is constructed. The multi-level spatial context of patches is also embedded in the graph, based on which image-level labels can be propagated to segments optimally. The contextual patch labeling process is formulated in an optimization framework and solved by a convergent iterative method. The category models are learned from the decomposed label representations of the image set and applied to the segments. Final labeling is obtained by combining all the information on pixel level. The effectiveness of the proposed method is demonstrated in experiments on two benchmark datasets and comparisons are taken. [ABSTRACT FROM AUTHOR]