WACV 2021 Papers

Skeleton-based action recognition has attracted research attentions in recent years. One common drawback in currently popular skeleton-based human action recognition methods is that the sparse skeleton information alone is not sufficient to fully characterize human motion. This limitation makes several existing methods incapable of correctly classifying action categories which exhibit only subtle motion differences. In this paper, we propose a novel framework for employing human pose skeleton and joint-centered light-weight information jointly in a two-stream graph convolutional network, namely, JOLO-GCN. Specifically, we use Joint-aligned optical Flow Patches (JFP) to capture the local subtle motion around each joint as the pivotal joint-centered visual information. Compared to the pure skeleton-based baseline, this hybrid scheme effectively boosts performance, while keeping the computational and memory overheads low. Experiments on the NTU RGB+D, NTU RGB+D 120, and the Kinetics-Skeleton dataset demonstrate clear accuracy improvements attained by the proposed method over the state-of-the-art skeleton-based methods.

Modern video streaming service companies offer millions of video-titles for its customers. A lot of these titles have repetitive introductory and recap parts in the beginning that customers have to manually skip in order to achieve an uninterrupted viewing experience. To avoid this unnecessary friction, some of the services have recently added "skip-intro" and "skip-recap" buttons to their video players before the intro and recap parts start. To efficiently scale this experience to their entire catalogs, it is important to automate the process of finding the intro and recap portions of titles. In this work, we pose intro and recap detection as a supervised sequence labeling problem and propose a novel end-to-end deep learning framework to this end. Specifically, we use CNNs to extract both visual and audio features from videos, and fuse these features using a B-LSTM in order to capture the various long and short term dependencies among different frame-features over time. Finally, we use a CRF to jointly optimize the sequence labeling for the intro and recap parts of the titles. We present a thorough empirical analysis of our model compared to several other deep learning based architectures and demonstrate the superior performance of our approach.

Recent works such as Mixup and Cutmix have demonstrated the effectiveness of augmenting training data for deep models. These methods generate new data by generally blending random image contents and mixing their labels proportionally. However, this strategy tends to produce unreasonable training samples for fine-grained recognition, leading to limited improvement. This is because mixing random image contents may potentially produce images containing destructed object structures. Further, as the category differences mainly reside in small part regions, mixing labels proportionally to the number of mixed pixels might result in label noisy problem. To augment more reasonable training data, we propose Intra-class Part Swapping (InPS) that produces new data by performing attention-guided content swapping on input pairs from the same class. Compared with previous approaches, InPS avoids introducing noisy labels and ensures a likely holistic structure of objects in generated images. We demonstrate InPS outperforms the most recent augmentation approaches in both fine-grained recognition and weakly object localization. Further, by simply incorporating the mid-level feature learning, our proposed method achieves state-of-the-art performance in the literature while maintaining the simplicity and inference efficiency. Our code is publicly available.

In this paper we propose a novel human-centered approach for detecting forgery in face images, using dynamic prototypes as a form of visual explanations. Currently, most state-of-the-art deepfake detections are based on black-box models that process videos frame-by-frame for inference, and few closely examine their temporal inconsistencies. However, the existence of such temporal artifacts within deepfake videos is key in detecting and explaining deepfakes to a supervising human. To this end, we propose Dynamic Prototype Network (DPNet) -- an interpretable and effective solution that utilizes dynamic representations (i.e., prototypes) to explain deepfake temporal artifacts. Extensive experimental results show that DPNet achieves competitive predictive performance, even on unseen testing datasets such as Google's DeepFakeDetection, DeeperForensics, and Celeb-DF, while providing easy referential explanations of deepfake dynamics. On top of DPNet's prototypical framework, we further formulate temporal logic specifications based on these dynamics to check our model's compliance to desired temporal behaviors, hence providing trustworthiness for such critical detection systems.

It is well known that human gaze carries significant information about visual attention. However, there are three main difficulties in incorporating the gaze data in an attention mechanism of deep neural networks: 1) the gaze fixation points are likely to have measurement errors due to blinking and rapid eye movements; 2) it is unclear when and how much the gaze data is correlated with visual attention; and 3) gaze data is not always available in many real-world situations. In this work, we introduce an effective probabilistic approach to integrate human gaze into spatiotemporal attention for egocentric activity recognition. Specifically, we represent the locations of gaze fixation points as structured discrete latent variables to model their uncertainties. In addition, we model the distribution of gaze fixations using a variational method. The gaze distribution is learned during the training process so that the ground-truth annotations of gaze locations are no longer needed in testing situations since they are predicted from the learned gaze distribution. The predicted gaze locations are used to provide informative attentional cues to improve the recognition performance. Our method outperforms all the previous state-of-the-art approaches on EGTEA, which is a large-scale dataset for egocentric activity recognition provided with gaze measurements. We also perform an ablation study and qualitative analysis to demonstrate that our attention mechanism is effective.

While Generative Adversarial Networks (GANs) are fundamental to many generative modelling applications, they suffer from numerous issues. In this work, we propose a principled framework to simultaneously mitigate two fundamental issues in GANs: catastrophic forgetting of the discriminator and mode collapse of the generator. We achieve this by employing for GANs a contrastive learning and mutual information maximization approach, and perform extensive analyses to understand sources of improvements. Our approach significantly stabilizes GAN training and improves GAN performance for image synthesis across five datasets under the same training and evaluation conditions against state-of-the-art works. In particular, compared to the state-of-the-art SSGAN, our approach does not suffer from poorer performance on image domains such as faces, and instead improves performance significantly. Our approach is simple to implement and practical: it involves only one auxiliary objective, has low computational cost, and performs robustly across a wide range of training settings and datasets without any hyperparameter tuning. For reproducibility, our code is available in the open-source GAN library, Mimicry.

Automatic check-out (ACO) emerges as an integral component in recent self-service retailing stores, which aims at automatically detecting and counting the randomly placed products upon a check-out platform. Existing data-driven counting works still have difficulties in generalizing to real-world retail product counting scenarios, since (1) real check-out images are hard to collect or cover all products and their possible layouts, (2) rapid updating of the product list leads to frequent and tedious re-training of the counting models. To overcome these obstacles, we contribute a practical automatic check-out framework tailored to real-world retail product counting scenarios, consisting of a photorealistic exemplar augmentation to generate physically reliable and photorealistic check-out images from canonical exemplars scanned for each product, and an incremental learning strategy to match the updating nature of the ACO system with much fewer training effort. Through comprehensive studies, we show that the proposed IncreACO serves as an effective framework on recent Retail Product Checkout (RPC) dataset, where the proposed photorealistic exemplar augmentation remarkably improves the counting performance against the state-of-the-art methods (77.15% v.s. 72.83% in counting accuracy), whilst the proposed incremental learning framework consistently extends the counting performance to new categories.

Video captioning is the task of predicting a semantic and syntactically correct sequence of words given some context video. The most successful methods for video captioning have a strong dependency on the effectiveness of semantic representations learned from visual models, but often produce syntactically incorrect sentences which harms their performance on standard datasets. In this paper, we address this limitation by considering syntactic representation learning as an essential component of video captioning. We construct a visual-syntactic embedding by mapping into a common vector space a visual representation, that depends only on the video, with a syntactic representation that depends only on Part-of-Speech (POS) tagging structures of the video description. We integrate this joint representation into an encoder-decoder architecture that we call Visual-Semantic-Syntactic Aligned Network (SemSynAN), which guides the decoder (text generation stage) by aligning temporal compositions of visual, semantic, and syntactic representations. We tested our proposed architecture obtaining state-of-the-art results on two widely used video captioning datasets: the Microsoft Video Description (MSVD) dataset and the Microsoft Research Video-to-Text (MSR-VTT) dataset.

Deep learning models such as Resnets have resulted in state-of-the-art accuracy in many computer vision problems. Neural ordinary differential equations (NODE) provides a continuous depth generalization of Resnets and overcome drawbacks of Resnet such as model selection and parameter complexity. Though NODE is more robust than Resnet, we find that NODE based architectures are still far away from providing robustness and uncertainty handling required for many computer vision problems. We propose novel NODE models which address these drawbacks. In particular, we propose Gaussian processes (GPs) to model the fully connected neural networks in NODE (NODE-GP) to improve robustness and uncertainty handling capabilities of NODE. The proposed model is flexible to accommodate different NODE architectures, and further improves the model selection capabilities in NODEs. We also find that numerical techniques play an important role in modelling NODE robustness, and propose to use different numerical techniques to improve NODE robustness. We demonstrate the superior robustness and uncertainty handling capabilities of proposed models on adversarial attacks and out-of-distribution experiments for the image classification tasks.

Point cloud semantic segmentation plays an essential role in autonomous driving, providing vital information about drivable surfaces and nearby objects that can aid higher level tasks such as path planning and collision avoidance. While current 3D semantic segmentation networks focus on convolutional architectures that perform great for well represented classes, they show a significant drop in performance for underrepresented classes that share similar geometric features. We propose a novel Detection Aware 3D Semantic Segmentation (DASS) framework that explicitly leverages localization features from an auxiliary 3D object detection task. By utilizing multitask training, the shared feature representation of the network is guided to be aware of per class detection features that aid tackling the differentiation of geometrically similar classes. We additionally provide a pipeline that uses DASS to generate high recall proposals for existing 2-stage detectors and demonstrate that the added supervisory signal can be used to improve 3D orientation estimation capabilities. Extensive experiments on both the SemanticKITTI and KITTI object datasets show that DASS can improve 3D semantic segmentation results of geometrically similar classes up to 37.8% IoU in image FOV while maintaining high precision bird's-eye view (BEV) detection results.

In this paper, we propose an approach to improve few-shot classification performance using a composite rotation based auxiliary task. Few-shot classification methods aim to produce neural networks that perform well for classes with a large number of training samples and classes with less number of training samples. They employ techniques to enable the network to produce highly discriminative features that are also very generic. Generally, the better the quality and generic-nature of the features produced by the network, the better is the performance of the network on few-shot learning. Our approach aims to train networks to produce such features by using a self-supervised auxiliary task. Our proposed composite rotation based auxiliary task performs rotation at two levels, i.e., rotation of patches inside the image (inner rotation) and rotation of the whole image (outer rotation) and assigns one out of 16 rotation classes to the modified image. We then simultaneously train for the composite rotation prediction task along with the original classification task, which forces the network to learn high-quality generic features that help improve the few-shot classification performance. We experimentally show that our approach performs better than existing few-shot learning methods on multiple benchmark datasets.

Whilst Generative Adversarial Networks (GANs) have gained a reputation as powerful generative models, they are notoriously difficult to train and suffer from instability in optimisation. Recent methods for tackling this drawback have typically approached it by inducing better behaviour on the discriminator component of the GAN; these include loss function modification, gradient regularisation and weight normalisation to create a discriminator that is well-behaved from a Lipschitz perspective. In this paper, we propose a novel and orthogonal contribution which modifies the architecture of a GAN. Our method embeds the powerful discriminating capabilities inherent in decision forests within the discriminator of a GAN. Empirically, we test the effectiveness of our approach on the CIFAR-10, Oxford Flowers and CUB Birds datasets. We show that our technique is easy to incorporate into existing GAN baselines and offers improvements on Frechet-Inception Distance (FID) scores by as high as 56.1% over several GAN baselines.

Recently there has been an interest in the potential of learning generative models from a single image, as opposed to from a large dataset. This task is of significance, as it means that generative models can be used in domains where collecting a large dataset is not feasible. However, training a model capable of generating realistic images from only a single sample is a difficult problem. In this work, we conduct a number of experiments to understand the challenges of training these methods and propose some best practices that we found allowed us to generate improved results over previous work. One key piece is that, unlike prior single image generation methods, we concurrently train several stages in a sequential multi-stage manner, allowing us to learn models with fewer stages of increasing image resolution. Compared to a recent state of the art baseline, our model is up to six times faster to train, has fewer parameters, and can better capture the global structure of images.

Weakly Supervised Object Localization (WSOL) methods generate both classification and localization results by learning from only image category labels. Previous methods usually utilize class activation map (CAM) to obtain target object regions. However, most of them only focus on improving foreground object parts in CAM, but ignore the important effect of its background contents. In this paper, we propose a confidence segmentation (ConfSeg) module that builds a confidence score for each pixel in CAM without introducing additional hyper-parameters. The generated sample-specific confidence mask is able to indicate the extent of determination for each pixel in CAM, and further supervises additional CAM extended from internal feature maps. Besides, we introduce Co-supervised Augmentation (CoAug) module to capture feature-level representation for foreground and background parts in CAM separately. Then a metric loss is applied at batch sample level to augment distinguish ability of our model, which helps a lot to localize more related object parts. Our final model, CSoA, combines the two modules and achieves superior performance, e.g. 37.69% and 48.81% Top-1 localization error on CUB-200 and ILSVRC datasets, respectively, which outperforms all previous methods and becomes the new state-of-the-art.

Images recorded during the lifetime of computer vision based systems undergo a wide range of illumination and environmental conditions affecting the reliability of previously trained machine learning models. Image normalization is hence a valuable preprocessing component to enhance the models' robustness. To this end, we introduce a new strategy for the cost function formulation of encoder-decoder networks to average out all the unimportant information in the input images (e.g. environmental features and illumination changes) to focus on the reconstruction of the salient features (e.g. class instances). Our method exploits the availability of identical sceneries under different illumination and environmental conditions for which we formulate a partially impossible reconstruction target: the input image will not convey enough information to reconstruct the target in its entirety. Its applicability is assessed on three publicly available datasets. We combine the triplet loss as a regularizer in the latent space representation and a nearest neighbour search to improve the generalization to unseen illuminations and class instances. The importance of the aforementioned post-processing is highlighted on an automotive application. To this end, we release a synthetic dataset of sceneries from three different passenger compartments where each scenery is rendered under ten different illumination and environmental conditions: https://sviro.kl.dfki.de

Deception is a common phenomenon in society, both in our private and professional lives. However, humans are notoriously bad at accurate deception detection. Based on the literature, human accuracy of distinguishing between lies and truthful statements is 54% on average, in other words, it is slightly better than a random guess. While people do not much care about this issue, in high-stakes situations such as interrogations for series crimes and for evaluating the testimonies in court cases, accurate deception detection methods are highly desirable. To achieve a reliable, covert, and non-invasive deception detection, we propose a novel method that disentangles facial expression and head pose related features using 2D-to-3D face reconstruction technique from a video sequence and uses them to learn characteristics of deceptive behavior. We evaluate the proposed method on the Real-Life Trial (RLT) dataset that contains high-stakes deceits recorded in courtrooms. Our results show that the proposed method (with an accuracy of 68%) improves the state of the art. Besides, a new dataset has been collected, for the first time, for low-stake deceit detection. In addition, we compare high-stake deceit detection methods on the newly collected low-stake deceits.

This paper proposes a novel Importance Guided Sparse Spatio-Temporal Regularization based Correlation Filter (IGSSTRCF) tracker. Our formulation explicitly models the variations in the correlation filters and associated spatial weights in successive frames. By imposing a sparsity penalty on these variations, the formulation ensures that only relevant changes are incorporated during updates. This results in more robust filter coefficients that minimize the tracking drift. The IGSSTRCF also includes an adaptive channel importance estimation strategy that assigns an importance weight to each feature channel during training. The proposed formulation is efficiently solved via the alternating direction method of multipliers. A comparative analysis is shown on TC128, UAV123, VOT-2017, and VOT-2019 datasets; and we present an ablation study to demonstrate the contribution of each component of the IGSSTRCF. It is observed that we outperform several state-of-the-art trackers and each component of the proposed IGSSTRCF contributes positively towards tracker performance.

Image inpainting is a non-trivial task in computer vi-sion due to multiple possibilities for filling the missing data, which may be dependent on the global information of the image. Most of the existing approaches use the attention mechanism to learn the global context of the image. This attention mechanism produces semantically plausible but blurry results because of incapability to capture the global context. In this paper, we introduce hypergraph convolution on spatial features to learn the complex relationship among the data. We introduce a trainable mechanism to connect nodes using hyperedges for hypergraph convolution. To the best of our knowledge, hypergraph convolution have never been used on spatial features for any image-to-image tasks in computer vision. Further, we introduce gated convolution in the discriminator to enforce local consistency in the predicted image. The experiments on Places2, CelebA-HQ, Paris Street View, and Facades datasets, show that our approach achieves state-of-the-art results.

Video-to-video translation is more difficult than image-to-image translation due to the temporal consistency problem that, if unaddressed, leads to distracting flickering effects. Although video models designed from scratch produce temporally consistent results, training them to match the vast visual knowledge captured by image models requires an intractable number of videos. To combine the benefits of image and video models, we propose an image-to-video model transfer method called Hyperconsistency (HyperCon) that transforms any well-trained image model into a temporally consistent video model without fine-tuning. HyperCon works by translating a temporally interpolated video frame-wise and then aggregating over temporally localized windows on the interpolated video. It handles both masked and unmasked inputs, enabling support for even more video-to-video translation tasks than prior image-to-video model transfer techniques. We demonstrate HyperCon on video style transfer and inpainting, where it performs favorably compared to prior state-of-the-art methods without training on a single stylized or incomplete video.