Recent successes of deep learning-based recognition rely on maintaining the content related to the main-task label. However, how to explicitly dispel the noisy signals for better generalization in a controllable manner remains an open issue. For instance, various factors such as identity-specific attributes, pose, illumination and expression affect the appearance of face images. Disentangling the identity-specific factors is potentially beneficial for facial expression recognition (FER). This chapter systematically summarize the detrimental factors as task-relevant/irrelevant semantic variations and unspecified latent variation. In this chapter, these problems are casted as either a deep metric learning problem or an adversarial minimax game in the latent space. For the former choice, a generalized adaptive (N+M)-tuplet clusters loss function together with the identity-aware hard-negative mining and online positive mining scheme can be used for identity-invariant FER. The better FER performance can be achieved by combining the deep metric loss and softmax loss in a unified two fully connected layer branches framework via joint optimization. For the latter solution, it is possible to equipping an end-to-end conditional adversarial network with the ability to decompose an input sample into three complementary parts. The discriminative representation inherits the desired invariance property guided by prior knowledge of the task, which is marginal independent to the task-relevant/irrelevant semantic and latent variations. The framework achieves top performance on a serial of tasks, including lighting, makeup, disguise-tolerant face recognition and facial attributes recognition. This chapter systematically summarize the popular and practical solution for disentanglement to achieve more discriminative visual recognition.
Forecasting pedestrian trajectories in dynamic scenes remains a critical problem with various applications, such as autonomous driving and socially aware robots. Such forecasting is challenging due to human-human and human-object interactions and future uncertainties caused by human randomness. Generative model-based methods handle future uncertainties by sampling a latent variable. However, few previous studies carefully explored the generation of the latent variable. In this work, we propose the Trajectory Predictor with Pseudo Oracle (TPPO), which is a generative model-based trajectory predictor. The first pseudo oracle is pedestrians' moving directions, and the second one is the latent variable estimated from observed trajectories. A social attention module is used to aggregate neighbors' interactions on the basis of the correlation between pedestrians' moving directions and their future trajectories. This correlation is inspired by the fact that a pedestrian's future trajectory is often influenced by pedestrians in front. A latent variable predictor is proposed to estimate latent variable distributions from observed and ground-truth trajectories. Moreover, the gap between these two distributions is minimized during training. Therefore, the latent variable predictor can estimate the latent variable from observed trajectories to approximate that estimated from ground-truth trajectories. We compare the performance of TPPO with related methods on several public datasets. Results demonstrate that TPPO outperforms state-of-the-art methods with low average and final displacement errors. Besides, the ablation study shows that the prediction performance will not dramatically decrease as sampling times decline during tests.
Pedestrian trajectory prediction in dynamic scenes remains a challenging and critical problem in numerous applications, such as self-driving cars and socially aware robots. Challenges concentrate on capturing pedestrians' social interactions and handling their future uncertainties. Pedestrians' head orientations can be used as an oracle that indicates relevant pedestrians[1], thus is beneficial to model social interactions. Moreover, latent variable distributions of pedestrians'future trajectories can be termed as another oracle. However, few works fully utilize these oracle information for an improved prediction performance. In this work, we propose GTPPO (Graph-based Trajectory Predictor with Pseudo Oracle), which is a generative model-based trajectory predictor. Pedestrians'social interactions are captured by the proposed GA2T (Graph Attention social Attention neTwork) module. Social attention is calculated on the basis of pedestrians' moving directions, which are termed as a pseudo oracle. Moreover, we propose a latent variable predictor to learn the latent variable distribution from observed trajectories. Such latent variable distribution reflects pedestrians'future trajectories, and therefore can be taken as another pseudo oracle. We compare the performance of GTPPO with several recently proposed methods on benchmarking datasets. Quantitative evaluations demonstrate that GTPPO outperforms state-of-the-art methods with lower average and final displacement errors. Qualitative evaluations show that GTPPO successfully recognizes the sudden motion changes since the estimated latent variable reflects the future trajectories.
We study the problem of learning disentangled representations for data across multiple domains and its applications in human retargeting. Our goal is to map an input image to an identity-invariant latent representation that captures intrinsic factors such as expressions and poses. To this end, we present a novel multi-view learning approach that leverages various data sources such as images, keypoints, and poses. Our model consists of multiple id-conditioned VAEs for different views of the data. During training, we encourage the latent embeddings to be consistent across these views. Our observation is that auxiliary data like keypoints and poses contain critical, id-agnostic semantic information, and it is easier to train a disentangling CVAE on these simpler views to separate such semantics from other id-specific attributes. We show that training multi-view CVAEs and encourage latent-consistency guides the image encoding to preserve the semantics of expressions and poses, leading to improved disentangled representations and better human retargeting results.
AI Safety is a major concern in many deep learning applications such as autonomous driving. Given a trained deep learning model, an important natural problem is how to reliably verify the model's prediction. In this paper, we propose a novel framework --- deep verifier networks (DVN) to verify the inputs and outputs of deep discriminative models with deep generative models. Our proposed model is based on conditional variational auto-encoders with disentanglement constraints. We give both intuitive and theoretical justifications of the model. Our verifier network is trained independently with the prediction model, which eliminates the need of retraining the verifier network for a new model. We test the verifier network on out-of-distribution detection and adversarial example detection problems, as well as anomaly detection problems in structured prediction tasks such as image caption generation. We achieve state-of-the-art results in all of these problems.
The labels in medical diagnosis task are usually discrete and successively distributed. For example, the Diabetic Retinopathy Diagnosis (DR) involves five health risk levels: no DR (0), mild DR (1), moderate DR (2), severe DR (3) and proliferative DR (4). This labeling system is common for medical disease. Previous methods usually construct a multi-binary-classification task or propose some re-parameter schemes in the output unit. In this paper, we target on this task from the perspective of loss function. More specifically, the Wasserstein distance is utilized as an alternative, explicitly incorporating the inter-class correlations by pre-defining its ground metric. Then, the ground metric which serves as a linear, convex or concave increasing function w.r.t. the Euclidean distance in a line is explored from an optimization perspective. Meanwhile, this paper also proposes of constructing the smoothed target labels that model the inlier and outlier noises by using a unimodal-uniform mixture distribution. Different from the one-hot setting, the smoothed label endues the computation of Wasserstein distance with more challenging features. With either one-hot or smoothed target label, this paper systematically concludes the practical closed-form solution. We evaluate our method on several medical diagnosis tasks (e.g., Diabetic Retinopathy and Ultrasound Breast dataset) and achieve state-of-the-art performance.
This paper targets the task with discrete and periodic class labels ($e.g.,$ pose/orientation estimation) in the context of deep learning. The commonly used cross-entropy or regression loss is not well matched to this problem as they ignore the periodic nature of the labels and the class similarity, or assume labels are continuous value. We propose to incorporate inter-class correlations in a Wasserstein training framework by pre-defining ($i.e.,$ using arc length of a circle) or adaptively learning the ground metric. We extend the ground metric as a linear, convex or concave increasing function $w.r.t.$ arc length from an optimization perspective. We also propose to construct the conservative target labels which model the inlier and outlier noises using a wrapped unimodal-uniform mixture distribution. Unlike the one-hot setting, the conservative label makes the computation of Wasserstein distance more challenging. We systematically conclude the practical closed-form solution of Wasserstein distance for pose data with either one-hot or conservative target label. We evaluate our method on head, body, vehicle and 3D object pose benchmarks with exhaustive ablation studies. The Wasserstein loss obtaining superior performance over the current methods, especially using convex mapping function for ground metric, conservative label, and closed-form solution.
Recent advances in domain adaptation show that deep self-training presents a powerful means for unsupervised domain adaptation. These methods often involve an iterative process of predicting on target domain and then taking the confident predictions as pseudo-labels for retraining. However, since pseudo-labels can be noisy, self-training can put overconfident label belief on wrong classes, leading to deviated solutions with propagated errors. To address the problem, we propose a confidence regularized self-training (CRST) framework, formulated as regularized self-training. Our method treats pseudo-labels as continuous latent variables jointly optimized via alternating optimization. We propose two types of confidence regularization: label regularization (LR) and model regularization (MR). CRST-LR generates soft pseudo-labels while CRST-MR encourages the smoothness on network output. Extensive experiments on image classification and semantic segmentation show that CRSTs outperform their non-regularized counterpart with state-of-the-art performance. The code and models of this work are available at https://github.com/yzou2/CRST.
This paper considers the problem of image set-based face verification and identification. Unlike traditional single sample (an image or a video) setting, this situation assumes the availability of a set of heterogeneous collection of orderless images and videos. The samples can be taken at different check points, different identity documents $etc$. The importance of each image is usually considered either equal or based on a quality assessment of that image independent of other images and/or videos in that image set. How to model the relationship of orderless images within a set remains a challenge. We address this problem by formulating it as a Markov Decision Process (MDP) in a latent space. Specifically, we first propose a dependency-aware attention control (DAC) network, which uses actor-critic reinforcement learning for attention decision of each image to exploit the correlations among the unordered images. An off-policy experience replay is introduced to speed up the learning process. Moreover, the DAC is combined with a temporal model for videos using divide and conquer strategies. We also introduce a pose-guided representation (PGR) scheme that can further boost the performance at extreme poses. We propose a parameter-free PGR without the need for training as well as a novel metric learning-based PGR for pose alignment without the need for pose detection in testing stage. Extensive evaluations on IJB-A/B/C, YTF, Celebrity-1000 datasets demonstrate that our method outperforms many state-of-art approaches on the set-based as well as video-based face recognition databases.
We consider the problem of comparing the similarity of image sets with variable-quantity, quality and un-ordered heterogeneous images. We use feature restructuring to exploit the correlations of both inner$\&$inter-set images. Specifically, the residual self-attention can effectively restructure the features using the other features within a set to emphasize the discriminative images and eliminate the redundancy. Then, a sparse/collaborative learning-based dependency-guided representation scheme reconstructs the probe features conditional to the gallery features in order to adaptively align the two sets. This enables our framework to be compatible with both verification and open-set identification. We show that the parametric self-attention network and non-parametric dictionary learning can be trained end-to-end by a unified alternative optimization scheme, and that the full framework is permutation-invariant. In the numerical experiments we conducted, our method achieves top performance on competitive image set/video-based face recognition and person re-identification benchmarks.