WACV 2023 Papers

Image completion is a research field which aims to generate visual contents for unknown regions of an image. Image outpainting and wide-range image blending, which we refer to as extensive painting, are considered challenging because compared to the large unknown regions, relatively less context is provided. Some recent studies have tried to decrease the complexity of extensive painting by generating image hints for the missing regions. In this paper, we introduce a novel modality of hints, the natural language. Moreover, we propose a Captioning-based Extensive Painting (CEP) module, which combines models for two different multi-modal tasks: image captioning and text-guided image completion. In order to generate appropriate captions for masked images, the image captioning model is optimized using self-critical sequence training (SCST) method with random masks. The biggest benefit of our methodology is the accessibility to well-designed image captioning and text-guided image manipulation models such as OFA and GLIDE without the need for additional architectural changes. In evaluation, our model demonstrates remarkable performance even with complicated image datasets both quantitatively and qualitatively.

A number of recent self-supervised learning methods have shown impressive performance on image classification and other tasks. A somewhat bewildering variety of techniques have been used, not always with a clear understanding of the reasons for their benefits, especially when used in combination. Here we treat the embeddings of images as point particles and consider model optimization as a dynamic process on this system of particles. Our dynamic model combines an attractive force for similar images, a locally dispersive force to avoid local collapse, and a global dispersive force to achieve a globally-homogeneous distribution of particles. The dynamic perspective highlights the advantage of using a delayed-parameter image embedding (a la BYOL) together with multiple views of the same image. It also uses a purely-dynamic local dispersive force (Brownian motion) that shows improved performance over other methods and does not require knowledge of other particle coordinates. The method is called MSBReg which stands for (i) a Multiview centroid loss, which applies an attractive force to pull different image view embeddings toward their centroid, (ii) a Singular value loss, which pushes the particle system toward spatially homogeneous density, (iii) a Brownian diffusive loss. We evaluate downstream classification performance of MSBReg on ImageNet as well as transfer learning tasks including fine-grained classification, multi-class object classification, object detection, and instance segmentation. In addition, we also show that applying our regularization term to other methods further improves their performance and stabilize the training by preventing a mode collapse.

Attention filtering under various learning scenarios has proven advantageous in enhancing the performance of many neural network architectures. The mainstream attention mechanism is established upon the non-local block, also known as an essential component of the prominent Transformer networks, to catch long-range correlations. However, such unilateral attention is often hampered by sparse and obscure responses, revealing insufficient dependencies across images/patches, and high computational cost, especially for those employing the multi-head design. To overcome these issues, we introduce a novel mechanism of aggregating bilateral attention (ABA) and validate its usefulness in tackling the task of few-shot instance localization, reflecting the underlying query-support dependency. Specifically, our method facilitates uncovering informative features via assessing: i) an embedding norm for exploring the semantically-related cues; ii) context awareness for correlating the query data and support regions. ABA is then carried out by integrating the affinity relations derived from the two measurements to serve as a lightweight but effective query-support attention mechanism with high localization recall. We evaluate ABA on two localization tasks, namely, few-shot action localization and one-shot object detection. Extensive experiments demonstrate that the proposed ABA achieves superior performances over existing methods.

Aerial imagery is widely utilized in visual data dependent applications such as military surveillance, earthquake assessment, etc. For these applications, minute texture in the aerial image are essential as any disturbance can cause inaccurate prediction. However, atmospheric haze severely reduces the visibility of the scene to be analysed, and hence takes a toll on accuracy of higher level applications. Existing methods either utilize additional prior while training, or produce sub-optimal outputs on different densities of haze degradation, due to absence of local and global dependencies in the extracted features. Therefore, it is essential to have a texture preserving algorithm for aerial image dehazing. In light of this, we propose a work that introduces a novel deformable multi-head attention with spatially attentive offset extraction based solution for aerial image dehazing. Here, the deformable multi-head attention is introduced to reconstruct fine level texture in the restored image. We also introduce spatially attentive offset extractor in the deformable convolution for focusing on relevant contextual information. Further, edge boosting skip connections are proposed for effectively passing edge features from shallow layers to deeper layers of the network. Thorough experimentation on synthetic as well as real-world data, along with extensive ablation study, demonstrate that the proposed method outperforms the prevailing works on aerial image dehazing. The code is provided at https://github.com/ AshutoshKulkarni4998/AIDTransformer.

Recent class-incremental learning methods combine deep neural architectures and learning algorithms to handle streaming data under memory and computational constraints. The performance of existing methods varies depending on the characteristics of the incremental process. To date, there is no other approach than to test all pairs of learning algorithms and neural architectures on the training data available at the start of the learning process to select a suited algorithm-architecture combination. To tackle this problem, in this article, we introduce AdvisIL, a method which takes as input the main characteristics of the incremental process (memory budget for the deep model, initial number of classes, size of incremental steps) and recommends an adapted pair of learning algorithm and neural architecture. The recommendation is based on a similarity between the user-provided settings and a large set of pre-computed experiments. AdvisIL makes class-incremental learning easier, since users do not need to run cumbersome experiments to design their system. We evaluate our method on four datasets under six incremental settings and three deep model sizes. We compare six algorithms and three deep neural architectures. Results show that AdvisIL has better overall performance than any of the individual combinations of a learning algorithm and a neural architecture. AdvisIL's code is available at https://github.com/EvaJF/AdvisIL.

Several works have shown that deep learning models are vulnerable to adversarial attacks where seemingly simple label-preserving changes to the input image lead to incorrect predictions. To combat this, gradient based adversarial training is generally employed as a standard defense mechanism. However, in cases where the loss landscape is discontinuous with respect to a given perturbation set, first order methods get stuck in local optima, and fail to defend against threat. This is often a problem for many physically realizable perturbation sets such as 2D affine transformations and 3D scene parameters. To work in such settings, we introduce a new optimization framework that alternates between global zeroth order sampling and local gradient updates to compute strong adversaries that can be used to harden the model against attack. Further, we design a powerful optimization algorithm using this framework, called Alternating Evolutionary Sampling and Descent (ASD), which combines an evolutionary search strategy (viz. covariance matrix adaptation) with gradient descent. We consider two settings with discontinuous/discrete and non-convex loss landscapes to evaluate ASD: a) 3D scene parameters and b) 2D patch attacks, and find that it achieves state-of-the-art results on adversarial robustness.

Knowledge distillation is a process of distilling information from a large model with significant knowledge capacity (teacher) to enhance a smaller model (student). Therefore, exploring the properties of the teacher is the key to improving student performance (e.g., teacher decision boundaries). One decision boundary exploring technique is to leverage adversarial attack methods, which add crafted perturbations within a ball constraint to clean inputs to create attack examples of the teacher called adversarial examples. These adversarial examples are informative examples because they are near decision boundaries. In this paper, we formulate a teacher adversarial local distribution, a set of all adversarial examples within the ball constraint given an input. This distribution is used to sufficiently explore the decision boundaries of the teacher by covering the full spectrum of possible teacher model perturbations. The student model is then regularized by matching the loss between teacher and student using these adversarial example inputs. We conducted a number of experiments on CIFAR-100 and Imagenet datasets to illustrate this teacher adversarial local distribution regularization (TALD) can be applied to improve performance of many existing knowledge distillation methods (e.g., KD, FitNet, CRD, VID, FT, etc.).

Contrastive learning is one of the fastest growing research areas in machine learning due to its ability to learn useful representations without labeled data. However, contrastive learning is susceptible to feature suppression - i.e., it may discard important information relevant to the task of interest, and learn irrelevant features. Past work has addressed this limitation via handcrafted data augmentations that eliminate irrelevant information. This approach however does not work across all datasets and tasks. Further, data augmentations fail in addressing feature suppression in multi-attribute classification when one attribute can suppress features relevant to other attributes. In this paper, we analyze the objective function of contrastive learning and formally prove that it is vulnerable to feature suppression. We then present Predictive Contrastive Learning (PrCL), a framework for learning unsupervised representations that are robust to feature suppression. The key idea is to force the learned representation to predict the input, and hence prevent it from discarding important information. Extensive experiments verify that PrCL is robust to feature suppression and outperforms state-of-the-art contrastive learning methods on a variety of datasets and tasks.

Generating images from hand-drawings is a crucial and fundamental task in content creation. The translation is difficult as there exist infinite possibilities and the different users usually expect different outcomes. Therefore, we propose a unified framework supporting a three-dimensional control over the image synthesis from sketches and strokes based on diffusion models. Users can not only decide the level of faithfulness to the input strokes and sketches, but also the degree of realism, as the user inputs are usually not consistent with the real images. Qualitative and quantitative experiments demonstrate that our framework achieves state-of-the-art performance while providing flexibility in generating customized images with control over shape, color, and realism. Moreover, our method unleashes applications such as editing on real images, generation with partial sketches and strokes, and multi-domain multi-modal synthesis.

Deep Neural Networks (DNNs) have been shown to be susceptible to memorization or overfitting in the presence of noisily-labelled data. For the problem of robust learning under such noisy data, several algorithms have been proposed. A prominent class of algorithms rely on sample selection strategies wherein, essentially, a fraction of samples with loss values below a certain threshold are selected for training. These algorithms are sensitive to such thresholds, and it is difficult to fix or learn these thresholds. Often, these algorithms also require information such as label noise rates which are typically unavailable in practice. In this paper, we propose an adaptive sample selection strategy that relies only on batch statistics of a given mini-batch to provide robustness against label noise. The algorithm does not have any additional hyperparameters for sample selection, does not need any information on noise rates and does not need access to separate data with clean labels. We empirically demonstrate the effectiveness of our algorithm on benchmark datasets.

Skeleton-based action recognition receives increasing attention because skeleton sequences reduce training complexity by eliminating visual information irrelevant to actions. To further improve sample efficiency, meta-learning-based one-shot learning solutions were developed for skeleton-based action recognition. These methods predict by finding the nearest neighbors according to the similarity between instance-level global embedding. However, such measurement holds unstable representativity due to inadequate generalized learning on the averaged local invariant and noisy features, while intuitively, steady and fine-grained recognition relies on determining key local body movements. To address this limitation, we present the Adaptive Local-Component-aware Graph Convolutional Network, which replaces the comparison metric with a focused sum of similarity measurements on aligned local embedding of action-critical spatial/temporal segments. Comprehensive one-shot experiments on the public benchmark of NTU-RGB+D 120 indicate that our method provides a stronger representation than the global embedding and helps our model reach state-of-the-art.

Cooperative perception allows a Connected Autonomous Vehicle (CAV) to interact with the other CAVs in the vicinity to enhance perception of surrounding objects to increase safety and reliability. It can compensate for the limitations of the conventional vehicular perception such as blind spots, low resolution, and weather effects. An effective feature fusion model for the intermediate fusion methods of cooperative perception can improve feature selection and information aggregation to further enhance the perception accuracy. We propose adaptive feature fusion models with trainable feature selection modules. One of our proposed models Spatial-wise Adaptive feature Fusion (S-AdaFusion) outperforms all other State-of-the-Arts (SOTAs) on two subsets of the OPV2V dataset: Default CARLA Towns for vehicle detection and the Culver City for domain adaptation. In addition, previous studies have only tested cooperative perception for vehicle detection. A pedestrian, however, is much more likely to be seriously injured in a traffic accident. We evaluate the performance of cooperative perception for both vehicle and pedestrian detection using the CODD dataset. Our architecture achieves higher Average Precision (AP) than other existing models for both vehicle and pedestrian detection on the CODD dataset. The experiments demonstrate that cooperative perception also improves the pedestrian detection accuracy compared to the conventional single vehicle perception process.

The stochastic gradient descent (SGD) optimizers are generally used to train the convolutional neural networks (CNNs). In recent years, several adaptive momentum based SGD optimizers have been introduced, such as Adam, diffGrad, Radam and AdaBelief. However, the existing SGD optimizers do not exploit the gradient norm of past iterations and lead to poor convergence and performance. In this paper, we propose a novel AdaNorm based SGD optimizers by correcting the norm of gradient in each iteration based on the adaptive training history of gradient norm. By doing so, the proposed optimizers are able to maintain high and representive gradient throughout the training and solves the low and atypical gradient problems. The proposed concept is generic and can be used with any existing SGD optimizer. We show the efficacy of the proposed AdaNorm with four state-of-the-art optimizers, including Adam, diffGrad, Radam and AdaBelief. We depict the performance improvement due to the proposed optimizers using three CNN models, including VGG16, ResNet18 and ResNet50, on three benchmark object recognition datasets, including CIFAR10, CIFAR100 and TinyImageNet.

Temporal Action Localization (TAL) is a significant and challenging task that searches for subtle human activities in an untrimmed video. To extract snippet-level video features, existing TAL methods commonly use video encoders pre-trained on short-video classification datasets. However, the snippet-level features can incur ambiguity between consecutive frames due to short and poor temporal information, disrupting the precise prediction of action instances. Several methods incorporating temporal relations have been proposed to mitigate this problem; however, they still suffer from poor video features. To address this issue, we propose a novel temporal action localization framework called an Action-aware Masking Network (AMNet). Our method simultaneously refines video features using action-aware attention and considers inherent temporal relations using self-attention and cross-attention mechanisms. First, we present an Action Masking Encoder (AME) that generates an action-aware mask to represent positive characteristics, which is then used to refine snippet-level features to be more salient around actions. Second, we design a Group Attention Module (GAM), which models relations of temporal information and exchanges mutual information by dividing the features into two groups, i.e., long and short-groups. Extensive experiments and ablation studies on two primary benchmark datasets demonstrate the effectiveness of AMNet, and our method achieves state-of-the-art performances on THUMOS-14 and ActivityNet1.3.

Vision Transformers has demonstrated competitive performance on computer vision tasks benefiting from their ability to capture long-range dependencies with multi-head self-attention modules and multi-layer perceptron. However, calculating global attention brings another disadvantage compared with convolutional neural networks, i.e. requiring much more data and computations to converge, which makes it difficult to generalize well on small datasets, which is common in practical applications. Previous works are either focusing on transferring knowledge from large datasets or adjusting the structure for small datasets. After carefully examining the self-attention modules, we discover that the number of trivial attention weights is far greater than the important ones and the accumulated trivial weights are dominating the attention in Vision Transformers due to their large quantity, which is not handled by the attention itself. This will cover useful non-trivial attention and harm the performance when trivial attention includes more noise, e.g. in shallow layers for some backbones. To solve this issue, we proposed to divide attention weights into trivial and non-trivial ones by thresholds, then Suppressing Accumulated Trivial Attention (SATA) weights by proposed Trivial WeIghts Suppression Transformation (TWIST) to reduce attention noise. Extensive experiments on CIFAR-100 and Tiny-ImageNet datasets show that our suppressing method boosts the accuracy of Vision Transformers by up to 2.3%. Code is available at https://github.com/xiangyu8/SATA.

Recently the focus of the computer vision community has shifted from expensive supervised learning towards self-supervised learning of visual representations. While the performance gap between supervised and self-supervised has been narrowing, the time for training self-supervised deep networks remains an order of magnitude larger than its supervised counterparts, which hinders progress, imposes carbon cost, and limits societal benefits to institutions with substantial resources. Motivated by these issues, this paper investigates reducing the training time of recent self-supervised methods by various model-agnostic strategies that have not been used for this problem. In particular, we study three strategies: an extendable cyclic learning rate schedule, a matching progressive augmentation magnitude and image resolutions schedule, and a hard positive mining strategy based on augmentation difficulty. We show that all three methods combined lead up to 2.7 times speed-up in the training time of several self-supervised methods while retaining comparable performance to the standard self-supervised learning setting.

An audio-visual event (AVE) is denoted by the correspondence of the visual and auditory signals in a video segment. Precise localization of the AVEs is very challenging since it demands effective multi-modal feature correspondence to ground the short and long range temporal interactions. Existing approaches struggle in capturing the different scales of multi-modal interaction due to ineffective multi-modal training strategies. To overcome this limitation, we introduce AVE-CLIP, a novel framework that integrates the AudioCLIP pre-trained on large-scale audio-visual data with a multi-window temporal transformer to effectively operate on different temporal scales of video frames. Our contributions are three-fold: (1) We introduce a multi-stage training framework to incorporate AudioCLIP pre-trained with audio-image pairs into the AVE localization task on video frames through contrastive fine-tuning, effective mean video feature extraction, and multi-scale training phases. (2) We propose a multi-domain attention mechanism that operates on both temporal and feature domains over varying timescales to fuse the local and global feature variations. (3) We introduce a temporal refining scheme with event-guided attention followed by a simple-yet-effective post processing step to handle significant variations of the background over diverse events. Our method achieves state-of-the-art performance on the publicly available AVE dataset with 5.9% mean accuracy improvement which proves its superiority over existing approaches.

Attention--or attribution--maps methods are methods designed to highlight regions of the model's input that were discriminative for its predictions. However, different attention maps methods can highlight different regions of the input, with sometimes contradictory explanations for a prediction. This effect is exacerbated when the training set is small. This indicates that either the model learned incorrect representations or that the attention maps methods did not accurately estimate the model's representations. We propose an unsupervised fine-tuning method that optimizes the consistency of attention maps and show that it improves both classification performance and the quality of attention maps. We propose an implementation for two state-of-the-art attention computation methods, Grad-CAM and Guided Backpropagation, which relies on an input masking technique. We also show results on Grad-CAM and Integrated Gradients in an ablation study. We evaluate this method on our own dataset of event detection in continuous video recordings of hospital patients aggregated and curated for this work. As a sanity check, we also evaluate the proposed method on PASCAL VOC and SVHN. With the proposed method, with small training sets, we achieve a 6.6 points lift of F1 score over the baselines on our video dataset, a 2.9 point lift of F1 score on PASCAL, and a 1.8 points lift of mean Intersection over Union over Grad-CAM for weakly supervised detection on PASCAL. Those improved attention maps may help clinicians better understand vision model predictions and ease the deployment of machine learning systems into clinical care. We share part of the code for this article at the following repository: https://github.com/alimirzazadeh/SemisupervisedAttention.

Although many long-range imaging systems are designed to support extended vision applications, a natural obstacle to their operation is degradation due to atmospheric turbulence. Atmospheric turbulence causes significant degradation to image quality by introducing blur and geometric distortion. In recent years, various deep learning-based single image atmospheric turbulence mitigation methods, including CNN-based and GAN inversion-based, have been proposed in the literature which attempt to remove the distortion in the image. However, some of these methods are difficult to train and often fail to reconstruct facial features and produce unrealistic results, especially in the case of high turbulence. Denoising Diffusion Probabilistic Models (DDPMs) have recently gained some traction because of their stable training process and their ability to generate high quality images. In this paper, we propose the first DDPM-based solution for the problem of atmospheric turbulence mitigation. We also propose a fast sampling technique for reducing the inference times for conditional DDPMs. Extensive experiments are conducted on synthetic and real-world data to show the significance of our model. To facilitate further research, all codes and pretrained models are publically available at http://github.com/Nithin-GK/AT-DDPM