Convolutional neural network-based approaches for semantic segmentation rely on supervision with pixel-level ground truth, but may not generalize well to unseen image domains. As the labeling process is tedious and labor intensive, developing algorithms that can adapt source truth labels target domain of great interest. In this paper, we propose an adversarial learning method adaptation in context segmentation. Considering segmentations as structured outputs contain spatial similarities...

We introduce a Deep Stochastic IOC RNN Encoder-decoder framework, DESIRE, for the task of future predictions multiple interacting agents in dynamic scenes. DESIRE effectively predicts locations objects scenes by 1) accounting multi-modal nature prediction (i.e., given same context, may vary), 2) foreseeing potential outcomes and make strategic based on that, 3) reasoning not only from past motion history, but also scene context as well interactions among agents. achieves these single...

This paper presents a novel large-scale dataset and comprehensive baselines for end-to-end pedestrian detection person recognition in raw video frames. Our address three issues: the performance of various combinations detectors recognizers, mechanisms to help improve overall re-identification (re-ID) accuracy assessing effectiveness different re-ID. We make distinct contributions. First, new dataset, PRW, is introduced evaluate Person Re-identification Wild, using videos acquired through six...

Despite recent advances in face recognition using deep learning, severe accuracy drops are observed for large pose variations unconstrained environments. Learning pose-invariant features is one solution, but needs expensively labeled large-scale data and carefully designed feature learning algorithms. In this work, we focus on frontalizing faces the wild under various head poses, including extreme profile view's. We propose a novel 3D Morphable Model (3DMM) conditioned Face Frontalization...

Predicting structured outputs such as semantic segmentation relies on expensive per-pixel annotations to learn supervised models like convolutional neural networks. However, trained one data domain may not generalize well other domains without for model finetuning. To avoid the labor-intensive process of annotation, we develop a adaptation method adapt source unlabeled target domain. We propose discriminative feature representations patches in by discovering multiple modes patch-wise output...

Despite the large volume of face recognition datasets, there is a significant portion subjects, which samples are insufficient and thus under-represented. Ignoring such results in training data. Training with under-represented data leads to biased classifiers conventionally-trained deep networks. In this paper, we propose center-based feature transfer framework augment space subjects from regular that have sufficiently diverse samples. A Gaussian prior variance assumed across all ones...

Reconstructing shape and reflectance properties from images is a highly under-constrained problem, has previously been addressed by using specialized hardware to capture calibrated data or assuming known (or constrained) reflectance. In contrast, we demonstrate that can recover non-Lambertian, spatially-varying BRDFs complex geometry belonging any arbitrary class, single RGB image captured under combination of unknown environment illumination flash lighting. We achieve this training deep...

We present a deep learning framework for accurate visual correspondences and demonstrate its effectiveness both geometric semantic matching, spanning across rigid motions to intra-class shape or appearance variations. In contrast previous CNN-based approaches that optimize surrogate patch similarity objective, we use metric directly learn feature space preserves either similarity. Our fully convolutional architecture, along with novel correspondence contrastive loss allows faster training by...

While several datasets for autonomous navigation have become available in recent years, they tended to focus on structured driving environments. This usually corresponds well-delineated infrastructure such as lanes, a small number of well-defined categories traffic participants, low variation object or background appearance and strong adherence rules. We propose DS, novel dataset road scene understanding unstructured environments where the above assumptions are largely not satisfied. It...

Large-scale training for semantic segmentation is challenging due to the expense of obtaining data this task relative other vision tasks. We propose a novel approach address difficulty. Given cheaply-obtained sparse image labelings, we propagate labels produce guessed dense labelings. A standard CNN-based network trained mimic these The label-propagation process defined via random-walk hitting probabilities, which leads differentiable parameterization with uncertainty estimates that are...

Data association problems are an important component of many computer vision applications, with multi-object tracking being one the most prominent examples. A typical approach to data involves finding a graph matching or network flow that minimizes sum pairwise costs, which often either hand-crafted learned as linear functions fixed features. In this work, we demonstrate it is possible learn features for network-flow-based via backpropagation, by expressing optimum smoothed problem...

We propose a deep inverse rendering framework for indoor scenes. From single RGB image of an arbitrary scene, we obtain complete scene reconstruction, estimating shape, spatially-varying lighting, and spatially-varying, non-Lambertian surface reflectance. Our novel network incorporates physical insights -- including spherical Gaussian lighting representation, differentiable layer to model appearance, cascade structure iteratively refine the predictions bilateral solver refinement allowing us...

Most modern frame interpolation approaches rely on explicit bidirectional optical flows between adjacent frames, thus are sensitive to the accuracy of underlying flow estimation in handling occlusions while additionally introducing computational bottlenecks unsuitable for efficient deployment. In this work, we propose a flow-free approach that is completely end-to-end trainable multi-frame video interpolation. Our method, FLAVR, leverages 3D spatio-temporal kernels directly learn motion...

Deep neural networks (DNNs) trained on large-scale datasets have recently achieved impressive improvements in face recognition. But a persistent challenge remains to develop methods capable of handling large pose variations that are relatively under-represented training data. This paper presents method for learning feature representation is invariant pose, without requiring extensive coverage We first propose generate non-frontal views from single frontal face, order increase the diversity...

We present an approach to matching images of objects in fine-grained datasets without using part annotations, with application the challenging problem weakly supervised single-view reconstruction. This is contrast prior works that require since across class and pose variations appearance features alone. overcome this challenge through a novel deep learning architecture, WarpNet, aligns object one image different another. exploit structure dataset create artificial data for training network...

Recognizing wild faces is extremely hard as they appear with all kinds of variations. Traditional methods either train specifically annotated variation data from target domains, or by introducing unlabeled to adapt the training data. Instead, we propose a universal representation learning framework that can deal larger unseen in given without leveraging domain knowledge. We firstly synthesize alongside some semantically meaningful variations, such low resolution, occlusion and head pose....

Scale drift is a crucial challenge for monocular autonomous driving to emulate the performance of stereo. This paper presents real-time SFM system that corrects scale using novel cue combination framework ground plane estimation, yielding accuracy comparable stereo over long sequences. Our estimation uses multiple cues like sparse features, dense inter-frame and (when applicable) object detection. A data-driven mechanism proposed learn models from training data relate observation covariances...

Supervised 3D reconstruction has witnessed a significant progress through the use of deep neural networks. However, this increase in performance requires large scale annotations 2D/3D data. In paper, we explore inexpensive 2D supervision as an alternative for expensive CAD annotation. Specifically, foreground masks weak raytrace pooling layer that enables perspective projection and backpropagation. Additionally, since from is ill posed problem, propose to constrain manifold unlabeled...

Despite rapid advances in face recognition, there remains a clear gap between the performance of still image-based recognition and video-based due to vast difference visual quality domains difficulty curating diverse large-scale video datasets. This paper addresses both those challenges, through an image feature-level domain adaptation approach, learn discriminative frame representations. The framework utilizes unlabeled data reduce different while transferring knowledge from labeled images....

In recent years, the need for semantic segmentation has arisen across several different applications and environments. However, expense redundancy of annotation often limits quantity labels available training in any domain, while deployment is easier if a single model works well domains. this paper, we pose novel problem universal semi-supervised propose solution framework, to meet dual needs lower costs. contrast counterpoints such as fine tuning, joint or unsupervised domain adaptation,...

We propose an active learning approach for transferring representations across domains. Our approach, adversarial domain adaptation (AADA), explores a duality between two related problems: alignment and importance sampling adapting models The former uses discriminative model to align domains, while the latter utilizes weigh samples account distribution shifts. Specifically, our weight promotes unlabeled with large uncertainty in classification diversity compared la-beled examples, thus...