The underlying correlation between audio and visual modalities within videos can be utilized to learn supervised information for unlabeled videos. In this paper, we present an end-to-end self-supervised framework named Audio-Visual Contrastive Learning (AVCL), to learn discriminative audio-visual representations for action recognition. Specifically, we design an attention based multi-modal fusion module (AMFM) to fuse audio and visual modalities. To align heterogeneous audio-visual modalities, we construct a novel co-correlation guided representation alignment module (CGRA). To learn supervised information from unlabeled videos, we propose a novel self-supervised contrastive learning module (SelfCL). Furthermore, to expand the existing audio-visual action recognition datasets and better evaluate our framework AVCL, we build a new audio-visual action recognition dataset named Kinetics-Sounds100. Experimental results on Kinetics-Sounds32 and Kinetics-Sounds100 datasets demonstrate the superiority of our AVCL over the state-of-the-art methods on large-scale action recognition benchmark.
Current Instance Transfer Learning (ITL) methodologies use domain adaptation and sub-space transformation to achieve successful transfer learning. However, these methodologies, in their processes, sometimes overfit on the target dataset or suffer from negative transfer if the test dataset has a high variance. Boosting methodologies have been shown to reduce the risk of overfitting by iteratively re-weighing instances with high-residual. However, this balance is usually achieved with parameter optimization, as well as reducing the skewness in weights produced due to the size of the source dataset. While the former can be achieved, the latter is more challenging and can lead to negative transfer. We introduce a simpler and more robust fix to this problem by building upon the popular boosting ITL regression methodology, two-stage TrAdaBoost.R2. Our methodology,~\us{}, is a boosting and random-forest based ensemble methodology that utilizes importance sampling to reduce the skewness due to the source dataset. We show that~\us{}~performs better than competitive transfer learning methodologies $63\%$ of the time. It also displays consistency in its performance over diverse datasets with varying complexities, as opposed to the sporadic results observed for other transfer learning methodologies.
Spatial-temporal representation learning is ubiquitous in various real-world applications, including visual comprehension, video understanding, multi-modal analysis, human-computer interaction, and urban computing. Due to the emergence of huge amounts of multi-modal heterogeneous spatial/temporal/spatial-temporal data in big data era, the existing visual methods rely heavily on large-scale data annotations and supervised learning to learn a powerful big model. However, the lack of interpretability, robustness, and out-of-distribution generalization are becoming the bottleneck problems of these models, which hinders the progress of interpretable and reliable artificial intelligence. The majority of the existing methods are based on correlation learning with the assumption that the data are independent and identically distributed, which lack an unified guidance and analysis about why modern spatial-temporal representation learning methods have limited interpretability and easily collapse into dataset bias. Inspired by the strong inference ability of human-level agents, recent years have therefore witnessed great effort in developing causal reasoning paradigms to realize robust representation and model learning with good interpretability. In this paper, we conduct a comprehensive review of existing causal reasoning methods for spatial-temporal representation learning, covering fundamental theories, models, and datasets. The limitations of current methods and datasets are also discussed. Moreover, we propose some primary challenges, opportunities, and future research directions for benchmarking causal reasoning algorithms in spatial-temporal representation learning.
Over parameterization is a common technique in deep learning to help models learn and generalize sufficiently to the given task; nonetheless, this often leads to enormous network structures and consumes considerable computing resources during training. Recent powerful transformer-based deep learning models on vision tasks usually have heavy parameters and bear training difficulty. However, many dense-prediction low-level computer vision tasks, such as rain streak removing, often need to be executed on devices with limited computing power and memory in practice. Hence, we introduce a recursive local window-based self-attention structure with residual connections and propose deraining a recursive transformer (DRT), which enjoys the superiority of the transformer but requires a small amount of computing resources. In particular, through recursive architecture, our proposed model uses only 1.3% of the number of parameters of the current best performing model in deraining while exceeding the state-of-the-art methods on the Rain100L benchmark by at least 0.33 dB. Ablation studies also investigate the impact of recursions on derain outcomes. Moreover, since the model contains no deliberate design for deraining, it can also be applied to other image restoration tasks. Our experiment shows that it can achieve competitive results on desnowing. The source code and pretrained model can be found at https://github.com/YC-Liang/DRT.
High-quality annotated images are significant to deep facial expression recognition (FER) methods. However, uncertain labels, mostly existing in large-scale public datasets, often mislead the training process. In this paper, we achieve uncertain label correction of facial expressions using auxiliary action unit (AU) graphs, called ULC-AG. Specifically, a weighted regularization module is introduced to highlight valid samples and suppress category imbalance in every batch. Based on the latent dependency between emotions and AUs, an auxiliary branch using graph convolutional layers is added to extract the semantic information from graph topologies. Finally, a re-labeling strategy corrects the ambiguous annotations by comparing their feature similarities with semantic templates. Experiments show that our ULC-AG achieves 89.31% and 61.57% accuracy on RAF-DB and AffectNet datasets, respectively, outperforming the baseline and state-of-the-art methods.
The neural network (NN) becomes one of the most heated type of models in various signal processing applications. However, NNs are extremely vulnerable to adversarial examples (AEs). To defend AEs, adversarial training (AT) is believed to be the most effective method while due to the intensive computation, AT is limited to be applied in most applications. In this paper, to resolve the problem, we design a generic and efficient AT improvement scheme, namely case-aware adversarial training (CAT). Specifically, the intuition stems from the fact that a very limited part of informative samples can contribute to most of model performance. Alternatively, if only the most informative AEs are used in AT, we can lower the computation complexity of AT significantly as maintaining the defense effect. To achieve this, CAT achieves two breakthroughs. First, a method to estimate the information degree of adversarial examples is proposed for AE filtering. Second, to further enrich the information that the NN can obtain from AEs, CAT involves a weight estimation and class-level balancing based sampling strategy to increase the diversity of AT at each iteration. Extensive experiments show that CAT is faster than vanilla AT by up to 3x while achieving competitive defense effect.
The lack of fine-grained 3D shape segmentation data is the main obstacle to developing learning-based 3D segmentation techniques. We propose an effective semi-supervised method for learning 3D segmentations from a few labeled 3D shapes and a large amount of unlabeled 3D data. For the unlabeled data, we present a novel multilevel consistency loss to enforce consistency of network predictions between perturbed copies of a 3D shape at multiple levels: point-level, part-level, and hierarchical level. For the labeled data, we develop a simple yet effective part substitution scheme to augment the labeled 3D shapes with more structural variations to enhance training. Our method has been extensively validated on the task of 3D object semantic segmentation on PartNet and ShapeNetPart, and indoor scene semantic segmentation on ScanNet. It exhibits superior performance to existing semi-supervised and unsupervised pre-training 3D approaches. Our code and trained models are publicly available at https://github.com/isunchy/semi_supervised_3d_segmentation.
Precise segmentation of teeth from intra-oral scanner images is an essential task in computer-aided orthodontic surgical planning. The state-of-the-art deep learning-based methods often simply concatenate the raw geometric attributes (i.e., coordinates and normal vectors) of mesh cells to train a single-stream network for automatic intra-oral scanner image segmentation. However, since different raw attributes reveal completely different geometric information, the naive concatenation of different raw attributes at the (low-level) input stage may bring unnecessary confusion in describing and differentiating between mesh cells, thus hampering the learning of high-level geometric representations for the segmentation task. To address this issue, we design a two-stream graph convolutional network (i.e., TSGCN), which can effectively handle inter-view confusion between different raw attributes to more effectively fuse their complementary information and learn discriminative multi-view geometric representations. Specifically, our TSGCN adopts two input-specific graph-learning streams to extract complementary high-level geometric representations from coordinates and normal vectors, respectively. Then, these single-view representations are further fused by a self-attention module to adaptively balance the contributions of different views in learning more discriminative multi-view representations for accurate and fully automatic tooth segmentation. We have evaluated our TSGCN on a real-patient dataset of dental (mesh) models acquired by 3D intraoral scanners. Experimental results show that our TSGCN significantly outperforms state-of-the-art methods in 3D tooth (surface) segmentation. Github: https://github.com/ZhangLingMing1/TSGCNet.
This paper addresses the color image completion problem in accordance with low-rank quatenrion matrix optimization that is characterized by sparse regularization in a transformed domain. This research was inspired by an appreciation of the fact that different signal types, including audio formats and images, possess structures that are inherently sparse in respect of their respective bases. Since color images can be processed as a whole in the quaternion domain, we depicted the sparsity of the color image in the quaternion discrete cosine transform (QDCT) domain. In addition, the representation of a low-rank structure that is intrinsic to the color image is a vital issue in the quaternion matrix completion problem. To achieve a more superior low-rank approximation, the quatenrion-based truncated nuclear norm (QTNN) is employed in the proposed model. Moreover, this model is facilitated by a competent alternating direction method of multipliers (ADMM) based on the algorithm. Extensive experimental results demonstrate that the proposed method can yield vastly superior completion performance in comparison with the state-of-the-art low-rank matrix/quaternion matrix approximation methods tested on color image recovery.
Intelligent reflecting surface (IRS) has been regarded as a promising and revolutionary technology for future wireless communication systems owing to its capability of tailoring signal propagation environment in an energy/spectrum/hardware-efficient manner. However, most existing studies on IRS optimizations are based on a simple and ideal reflection model that is impractical in hardware implementation, which thus leads to severe performance loss in realistic wideband/multi-band systems. To deal with this problem, in this paper we first propose a more practical and more tractable IRS reflection model that describes the difference of reflection responses for signals at different frequencies. Then, we investigate the joint transmit beamforming and IRS reflection beamforming design for an IRS-assisted multi-cell multi-band system. Both power minimization and sum-rate maximization problems are solved by exploiting popular second-order cone programming (SOCP), Riemannian manifold, minimization-majorization (MM), weighted minimum mean square error (WMMSE), and block coordinate descent (BCD) methods. Simulation results illustrate the significant performance improvement of our proposed joint transmit beamforming and reflection design algorithms based on the practical reflection model in terms of power saving and rate enhancement.