Computer Vision – ECCV 2018

Boosting is a classic algorithm which has been successfully applied to diverse computer vision tasks. In the scenario of image denoising, however, the existing boosting algorithms are surpassed by the emerging learning-based models. In this paper, we propose a novel deep boosting framework (DBF) for denoising, which integrates several convolutional networks in a feed-forward fashion. Along with the integrated networks, however, the depth of the boosting framework is substantially increased, which brings difficulty to training. To solve this problem, we introduce the concept of dense connection that overcomes the vanishing of gradients during training. Furthermore, we propose a path-widening fusion scheme cooperated with the dilated convolution to derive a lightweight yet efficient convolutional network as the boosting unit, named Dilated Dense Fusion Network (DDFN). Comprehensive experiments demonstrate that our DBF outperforms existing methods on widely used benchmarks, in terms of different denoising tasks.

As an agent moves through the world, the apparent motion of scene elements is (usually) inversely proportional to their depth (Strictly speaking, this statement is true only after one has compensated for camera rotation, individual object motion, and image position. We address these issues in the paper). It is natural for a learning agent to associate image patterns with the magnitude of their displacement over time: as the agent moves, faraway mountains don’t move much; nearby trees move a lot. This natural relationship between the appearance of objects and their motion is a rich source of information about the world. In this work, we start by training a deep network, using fully automatic supervision, to predict relative scene depth from single images. The relative depth training images are automatically derived from simple videos of cars moving through a scene, using recent motion segmentation techniques, and no human-provided labels. The proxy task of predicting relative depth from a single image induces features in the network that result in large improvements in a set of downstream tasks including semantic segmentation, joint road segmentation and car detection, and monocular (absolute) depth estimation, over a network trained from scratch. The improvement on the semantic segmentation task is greater than that produced by any other automatically supervised methods. Moreover, for monocular depth estimation, our unsupervised pre-training method even outperforms supervised pre-training with ImageNet. In addition, we demonstrate benefits from learning to predict (again, completely unsupervised) relative depth in the specific videos associated with various downstream tasks (e.g., KITTI). We adapt to the specific scenes in those tasks in an unsupervised manner to improve performance. In summary, for semantic segmentation, we present state-of-the-art results among methods that do not use supervised pre-training, and we even exceed the performance of supervised ImageNet pre-trained models for monocular depth estimation, achieving results that are comparable with state-of-the-art methods.

This work extends popular star-convexity and other more general forms of convexity priors. We represent an object as a union of “convex” overlappable subsets. Since an arbitrary shape can always be divided into convex parts, our regularization model restricts the number of such parts. Previous k-part shape priors are limited to disjoint parts. For example, one approach segments an object via optimizing its k coverage by disjoint convex parts, which we show is highly sensitive to local minima. In contrast, our shape model allows the convex parts to overlap, which both relaxes and simplifies the coverage problem, e.g. fewer parts are needed to represent any object. As shown in the paper, for many forms of convexity our regularization model is significantly more descriptive for any given k. Our shape prior is useful in practice, e.g. biomedical applications, and its optimization is robust to local minima.

We propose an end-to-end deep learning architecture that produces a 3D shape in triangular mesh from a single color image. Limited by the nature of deep neural network, previous methods usually represent a 3D shape in volume or point cloud, and it is non-trivial to convert them to the more ready-to-use mesh model. Unlike the existing methods, our network represents 3D mesh in a graph-based convolutional neural network and produces correct geometry by progressively deforming an ellipsoid, leveraging perceptual features extracted from the input image. We adopt a coarse-to-fine strategy to make the whole deformation procedure stable, and define various of mesh related losses to capture properties of different levels to guarantee visually appealing and physically accurate 3D geometry. Extensive experiments show that our method not only qualitatively produces mesh model with better details, but also achieves higher 3D shape estimation accuracy compared to the state-of-the-art.

Visual attention has shown usefulness in image captioning, with the goal of enabling a caption model to selectively focus on regions of interest. Existing models typically rely on top-down language information and learn attention implicitly by optimizing the captioning objectives. While somewhat effective, the learned top-down attention can fail to focus on correct regions of interest without direct supervision of attention. Inspired by the human visual system which is driven by not only the task-specific top-down signals but also the visual stimuli, we in this work propose to use both types of attention for image captioning. In particular, we highlight the complementary nature of the two types of attention and develop a model (Boosted Attention) to integrate them for image captioning. We validate the proposed approach with state-of-the-art performance across various evaluation metrics.

Existing deep learning based image inpainting methods use a standard convolutional network over the corrupted image, using convolutional filter responses conditioned on both valid pixels as well as the substitute values in the masked holes (typically the mean value). This often leads to artifacts such as color discrepancy and blurriness. Post-processing is usually used to reduce such artifacts, but are expensive and may fail. We propose the use of partial convolutions, where the convolution is masked and renormalized to be conditioned on only valid pixels. We further include a mechanism to automatically generate an updated mask for the next layer as part of the forward pass. Our model outperforms other methods for irregular masks. We show qualitative and quantitative comparisons with other methods to validate our approach.

Advances in photo editing and manipulation tools have made it significantly easier to create fake imagery. Learning to detect such manipulations, however, remains a challenging problem due to the lack of sufficient amounts of manipulated training data. In this paper, we propose a learning algorithm for detecting visual image manipulations that is trained only using a large dataset of real photographs. The algorithm uses the automatically recorded photo EXIF metadata as supervisory signal for training a model to determine whether an image is self-consistent — that is, whether its content could have been produced by a single imaging pipeline. We apply this self-consistency model to the task of detecting and localizing image splices. The proposed method obtains state-of-the-art performance on several image forensics benchmarks, despite never seeing any manipulated images at training. That said, it is merely a step in the long quest for a truly general purpose visual forensics tool.

Estimating the 3D pose of a hand is an essential part of human-computer interaction. Estimating 3D pose using depth or multi-view sensors has become easier with recent advances in computer vision, however, regressing pose from a single RGB image is much less straightforward. The main difficulty arises from the fact that 3D pose requires some form of depth estimates, which are ambiguous given only an RGB image. In this paper we propose a new method for 3D hand pose estimation from a monocular image through a novel 2.5D pose representation. Our new representation estimates pose up to a scaling factor, which can be estimated additionally if a prior of the hand size is given. We implicitly learn depth maps and heatmap distributions with a novel CNN architecture. Our system achieves state-of-the-art accuracy for 2D and 3D hand pose estimation on several challenging datasets in presence of severe occlusions.

Convolutional neural networks (CNN) are limited by the lack of capability to handle geometric information due to the fixed grid kernel structure. The availability of depth data enables progress in RGB-D semantic segmentation with CNNs. State-of-the-art methods either use depth as additional images or process spatial information in 3D volumes or point clouds. These methods suffer from high computation and memory cost. To address these issues, we present Depth-aware CNN by introducing two intuitive, flexible and effective operations: depth-aware convolution and depth-aware average pooling. By leveraging depth similarity between pixels in the process of information propagation, geometry is seamlessly incorporated into CNN. Without introducing any additional parameters, both operators can be easily integrated into existing CNNs. Extensive experiments and ablation studies on challenging RGB-D semantic segmentation benchmarks validate the effectiveness and flexibility of our approach.

We present an interpretable framework for path prediction that leverages dependencies between agents’ behaviors and their spatial navigation environment. We exploit two sources of information: the past motion trajectory of the agent of interest and a wide top-view image of the navigation scene. We propose a Clairvoyant Attentive Recurrent Network (CAR-Net) that learns where to look in a large image of the scene when solving the path prediction task. Our method can attend to any area, or combination of areas, within the raw image (e.g., road intersections) when predicting the trajectory of the agent. This allows us to visualize fine-grained semantic elements of navigation scenes that influence the prediction of trajectories. To study the impact of space on agents’ trajectories, we build a new dataset made of top-view images of hundreds of scenes (Formula One racing tracks) where agents’ behaviors are heavily influenced by known areas in the images (e.g., upcoming turns). CAR-Net successfully attends to these salient regions. Additionally, CAR-Net reaches state-of-the-art accuracy on the standard trajectory forecasting benchmark, Stanford Drone Dataset (SDD). Finally, we show CAR-Net’s ability to generalize to unseen scenes.

Inspired by the pioneering work of information bottleneck principle for Deep Neural Networks (DNNs) analysis, we design an information plane based framework to evaluate the capability of DNNs for image classification tasks, which not only helps understand the capability of DNNs, but also helps us choose a neural network which leads to higher classification accuracy more efficiently. Further, with experiments, the relationship among the model accuracy, I(X; T) and I(T; Y) are analyzed, where I(X; T) and I(T; Y) are the mutual information of DNN’s output T with input X and label Y. We also show the information plane is more informative than loss curve and apply mutual information to infer the model’s capability for recognizing objects of each class. Our studies would facilitate a better understanding of DNNs.

Face hallucination is a generative task to super-resolve the facial image with low resolution while human perception of face heavily relies on identity information. However, previous face hallucination approaches largely ignore facial identity recovery. This paper proposes Super-Identity Convolutional Neural Network (SICNN) to recover identity information for generating faces closed to the real identity. Specifically, we define a super-identity loss to measure the identity difference between a hallucinated face and its corresponding high-resolution face within the hypersphere identity metric space. However, directly using this loss will lead to a Dynamic Domain Divergence problem, which is caused by the large margin between the high-resolution domain and the hallucination domain. To overcome this challenge, we present a domain-integrated training approach by constructing a robust identity metric for faces from these two domains. Extensive experimental evaluations demonstrate that the proposed SICNN achieves superior visual quality over the state-of-the-art methods on a challenging task to super-resolve 12 $$\times $$ × 14 faces with an 8 $$\times $$ × upscaling factor. In addition, SICNN significantly improves the recognizability of ultra-low-resolution faces.

Despite tremendous progress achieved in temporal action localization, state-of-the-art methods still struggle to train accurate models when annotated data is scarce. In this paper, we introduce a novel active learning framework for temporal localization that aims to mitigate this data dependency issue. We equip our framework with active selection functions that can reuse knowledge from previously annotated datasets. We study the performance of two state-of-the-art active selection functions as well as two widely used active learning baselines. To validate the effectiveness of each one of these selection functions, we conduct simulated experiments on ActivityNet. We find that using previously acquired knowledge as a bootstrapping source is crucial for active learners aiming to localize actions. When equipped with the right selection function, our proposed framework exhibits significantly better performance than standard active learning strategies, such as uncertainty sampling. Finally, we employ our framework to augment the newly compiled Kinetics action dataset with ground-truth temporal annotations. As a result, we collect Kinetics-Localization, a novel large-scale dataset for temporal action localization, which contains more than 15K YouTube videos.

We propose a novel end-to-end semi-supervised adversarial framework to generate photorealistic face images of new identities with a wide range of expressions, poses, and illuminations conditioned by synthetic images sampled from a 3D morphable model. Previous adversarial style-transfer methods either supervise their networks with a large volume of paired data or train highly under-constrained two-way generative networks in an unsupervised fashion. We propose a semi-supervised adversarial learning framework to constrain the two-way networks by a small number of paired real and synthetic images, along with a large volume of unpaired data. A set-based loss is also proposed to preserve identity coherence of generated images. Qualitative results show that generated face images of new identities contain pose, lighting and expression diversity. They are also highly constrained by the synthetic input images while adding photorealism and retaining identity information. We combine face images generated by the proposed method with a real data set to train face recognition algorithms and evaluate the model quantitatively on two challenging data sets: LFW and IJB-A. The generated images by our framework consistently improve the performance of deep face recognition networks trained with the Oxford VGG Face dataset, and achieve comparable results to the state-of-the-art.

We introduce a deep learning-based method to generate full 3D hair geometry from an unconstrained image. Our method can recover local strand details and has real-time performance. State-of-the-art hair modeling techniques rely on large hairstyle collections for nearest neighbor retrieval and then perform ad-hoc refinement. Our deep learning approach, in contrast, is highly efficient in storage and can run 1000 times faster while generating hair with 30K strands. The convolutional neural network takes the 2D orientation field of a hair image as input and generates strand features that are evenly distributed on the parameterized 2D scalp. We introduce a collision loss to synthesize more plausible hairstyles, and the visibility of each strand is also used as a weight term to improve the reconstruction accuracy. The encoder-decoder architecture of our network naturally provides a compact and continuous representation for hairstyles, which allows us to interpolate naturally between hairstyles. We use a large set of rendered synthetic hair models to train our network. Our method scales to real images because an intermediate 2D orientation field, automatically calculated from the real image, factors out the difference between synthetic and real hairs. We demonstrate the effectiveness and robustness of our method on a wide range of challenging real Internet pictures, and show reconstructed hair sequences from videos.

A capsule is a collection of neurons which represents different variants of a pattern in the network. The routing scheme ensures only certain capsules which resemble lower counterparts in the higher layer should be activated. However, the computational complexity becomes a bottleneck for scaling up to larger networks, as lower capsules need to correspond to each and every higher capsule. To resolve this limitation, we approximate the routing process with two branches: a master branch which collects primary information from its direct contact in the lower layer and an aide branch that replenishes master based on pattern variants encoded in other lower capsules. Compared with previous iterative and unsupervised routing scheme, these two branches are communicated in a fast, supervised and one-time pass fashion. The complexity and runtime of the model are therefore decreased by a large margin. Motivated by the routing to make higher capsule have agreement with lower capsule, we extend the mechanism as a compensation for the rapid loss of information in nearby layers. We devise a feedback agreement unit to send back higher capsules as feedback. It could be regarded as an additional regularization to the network. The feedback agreement is achieved by comparing the optimal transport divergence between two distributions (lower and higher capsules). Such an add-on witnesses a unanimous gain in both capsule and vanilla networks. Our proposed EncapNet performs favorably better against previous state-of-the-arts on CIFAR10/100, SVHN and a subset of ImageNet.

Knowledge Transfer (KT) techniques tackle the problem of transferring the knowledge from a large and complex neural network into a smaller and faster one. However, existing KT methods are tailored towards classification tasks and they cannot be used efficiently for other representation learning tasks. In this paper we propose a novel probabilistic knowledge transfer method that works by matching the probability distribution of the data in the feature space instead of their actual representation. Apart from outperforming existing KT techniques, the proposed method allows for overcoming several of their limitations providing new insight into KT as well as novel KT applications, ranging from KT from handcrafted feature extractors to cross-modal KT from the textual modality into the representation extracted from the visual modality of the data.

Videos recorded from first person (egocentric) perspective have little visual appearance in common with those from third person perspective, especially with videos captured by top-view surveillance cameras. In this paper, we aim to relate these two sources of information from a surveillance standpoint, namely in terms of identification and temporal alignment. Given an egocentric video and a top-view video, our goals are to: (a) identify the egocentric camera holder in the top-view video (self-identification), (b) identify the humans visible in the content of the egocentric video, within the content of the top-view video (re-identification), and (c) temporally align the two videos. The main challenge is that each of these tasks is highly dependent on the other two. We propose a unified framework to jointly solve all three problems. We evaluate the efficacy of the proposed approach on a publicly available dataset containing a variety of videos recorded in different scenarios.

We present a method to populate an unknown environment with models of previously seen objects, placed in a Euclidean reference frame that is inferred causally and on-line using monocular video along with inertial sensors. The system we implement returns a sparse point cloud for the regions of the scene that are visible but not recognized as a previously seen object, and a detailed object model and its pose in the Euclidean frame otherwise. The system includes bottom-up and top-down components, whereby deep networks trained for detection provide likelihood scores for object hypotheses provided by a nonlinear filter, whose state serves as memory. Additional networks provide likelihood scores for edges, which complements detection networks trained to be invariant to small deformations. We test our algorithm on existing datasets, and also introduce the VISMA dataset, that provides ground truth pose, point-cloud map, and object models, along with time-stamped inertial measurements.

Current state-of-the-art approaches for spatio-temporal action localization rely on detections at the frame level and model temporal context with 3D ConvNets. Here, we go one step further and model spatio-temporal relations to capture the interactions between human actors, relevant objects and scene elements essential to differentiate similar human actions. Our approach is weakly supervised and mines the relevant elements automatically with an actor-centric relational network (ACRN). ACRN computes and accumulates pair-wise relation information from actor and global scene features, and generates relation features for action classification. It is implemented as neural networks and can be trained jointly with an existing action detection system. We show that ACRN outperforms alternative approaches which capture relation information, and that the proposed framework improves upon the state-of-the-art performance on JHMDB and AVA. A visualization of the learned relation features confirms that our approach is able to attend to the relevant relations for each action.

Humans have the amazing ability to perform very subtle manipulation task using a closed-loop control system with imprecise mechanics (i.e., our body parts) but rich sensory information (e.g., vision, tactile, etc.). In the closed-loop system, the ability to monitor the state of the task via rich sensory information is important but often less studied. In this work, we take liquid pouring as a concrete example and aim at learning to continuously monitor whether liquid pouring is successful (e.g., no spilling) or not via rich sensory inputs. We mimic humans’ rich sensories using synchronized observation from a chest-mounted camera and a wrist-mounted IMU sensor. Given many success and failure demonstrations of liquid pouring, we train a hierarchical LSTM with late fusion for monitoring. To improve the robustness of the system, we propose two auxiliary tasks during training: inferring (1) the initial state of containers and (2) forecasting the one-step future 3D trajectory of the hand with an adversarial training procedure. These tasks encourage our method to learn representation sensitive to container states and how objects are manipulated in 3D. With these novel components, our method achieves $$\sim $$ ∼ 8% and $$\sim $$ ∼ 11% better monitoring accuracy than the baseline method without auxiliary tasks on unseen containers and unseen users respectively.

The Convolutional Neural Network (CNN) based region proposal generation method (i.e. region proposal network), trained using bounding box annotations, is an essential component in modern fully supervised object detectors. However, Weakly Supervised Object Detection (WSOD) has not benefited from CNN-based proposal generation due to the absence of bounding box annotations, and is relying on standard proposal generation methods such as selective search. In this paper, we propose a weakly supervised region proposal network which is trained using only image-level annotations. The weakly supervised region proposal network consists of two stages. The first stage evaluates the objectness scores of sliding window boxes by exploiting the low-level information in CNN and the second stage refines the proposals from the first stage using a region-based CNN classifier. Our proposed region proposal network is suitable for WSOD, can be plugged into a WSOD network easily, and can share its convolutional computations with the WSOD network. Experiments on the PASCAL VOC and ImageNet detection datasets show that our method achieves the state-of-the-art performance for WSOD with performance gain of about $$3\%$$ 3 % on average.

Object detection is one of the major problems in computer vision, and has been extensively studied. Most of the existing detection works rely on labor-intensive supervision, such as ground truth bounding boxes of objects or at least image-level annotations. On the contrary, we propose an object detection method that does not require any form of human annotation on target tasks, by exploiting freely available web images. In order to facilitate effective knowledge transfer from web images, we introduce a multi-instance multi-label domain adaption learning framework with two key innovations. First of all, we propose an instance-level adversarial domain adaptation network with attention on foreground objects to transfer the object appearances from web domain to target domain. Second, to preserve the class-specific semantic structure of transferred object features, we propose a simultaneous transfer mechanism to transfer the supervision across domains through pseudo strong label generation. With our end-to-end framework that simultaneously learns a weakly supervised detector and transfers knowledge across domains, we achieved significant improvements over baseline methods on the benchmark datasets.

Current top-performing object detectors depend on deep CNN backbones, such as ResNet-101 and Inception, benefiting from their powerful feature representations but suffering from high computational costs. Conversely, some lightweight model based detectors fulfil real time processing, while their accuracies are often criticized. In this paper, we explore an alternative to build a fast and accurate detector by strengthening lightweight features using a hand-crafted mechanism. Inspired by the structure of Receptive Fields (RFs) in human visual systems, we propose a novel RF Block (RFB) module, which takes the relationship between the size and eccentricity of RFs into account, to enhance the feature discriminability and robustness. We further assemble RFB to the top of SSD, constructing the RFB Net detector. To evaluate its effectiveness, experiments are conducted on two major benchmarks and the results show that RFB Net is able to reach the performance of advanced very deep detectors while keeping the real-time speed. Code is available at https://github.com/ruinmessi/RFBNet .

In this paper, we make two contributions to unsupervised domain adaptation (UDA) using the convolutional neural network (CNN). First, our approach transfers knowledge in all the convolutional layers through attention alignment. Most previous methods align high-level representations, e.g., activations of the fully connected (FC) layers. In these methods, however, the convolutional layers which underpin critical low-level domain knowledge cannot be updated directly towards reducing domain discrepancy. Specifically, we assume that the discriminative regions in an image are relatively invariant to image style changes. Based on this assumption, we propose an attention alignment scheme on all the target convolutional layers to uncover the knowledge shared by the source domain. Second, we estimate the posterior label distribution of the unlabeled data for target network training. Previous methods, which iteratively update the pseudo labels by the target network and refine the target network by the updated pseudo labels, are vulnerable to label estimation errors. Instead, our approach uses category distribution to calculate the cross-entropy loss for training, thereby ameliorating the error accumulation of the estimated labels. The two contributions allow our approach to outperform the state-of-the-art methods by +2.6% on the Office-31 dataset.

In this paper, we present MultiPoseNet, a novel bottom-up multi-person pose estimation architecture that combines a multi-task model with a novel assignment method. MultiPoseNet can jointly handle person detection, person segmentation and pose estimation problems. The novel assignment method is implemented by the Pose Residual Network (PRN) which receives keypoint and person detections, and produces accurate poses by assigning keypoints to person instances. On the COCO keypoints dataset, our pose estimation method outperforms all previous bottom-up methods both in accuracy (+4-point mAP over previous best result) and speed; it also performs on par with the best top-down methods while being at least 4x faster. Our method is the fastest real time system with $$\sim 23$$ ∼ 23 frames/sec.

This work provides a simple approach to discover tight object bounding boxes with only image-level supervision, called Tight box mining with Surrounding Segmentation Context (TS2C). We observe that object candidates mined through current multiple instance learning methods are usually trapped to discriminative object parts, rather than the entire object. TS2C leverages surrounding segmentation context derived from weakly-supervised segmentation to suppress such low-quality distracting candidates and boost the high-quality ones. Specifically, TS2C is developed based on two key properties of desirable bounding boxes: (1) high purity, meaning most pixels in the box are with high object response, and (2) high completeness, meaning the box covers high object response pixels comprehensively. With such novel and computable criteria, more tight candidates can be discovered for learning a better object detector. With TS2C, we obtain 48.0% and 44.4% mAP scores on VOC 2007 and 2012 benchmarks, which are the new state-of-the-arts.

We introduce a method for improving convolutional neural networks (CNNs) for scene classification. We present a hierarchy of specialist networks, which disentangles the intra-class variation and inter-class similarity in a coarse to fine manner. Our key insight is that each subset within a class is often associated with different types of inter-class similarity. This suggests that existing network of experts approaches that organize classes into coarse categories are suboptimal. In contrast, we group images based on high-level appearance features rather than their class membership and dedicate a specialist model per group. In addition, we propose an alternating architecture with a global ordered- and a global orderless-representation to account for both the coarse layout of the scene and the transient objects. We demonstrate that it leads to better performance than using a single type of representation as well as the fused features. We also introduce a mini-batch soft k-means that allows end-to-end fine-tuning, as well as a novel routing function for assigning images to specialists. Experimental results show that the proposed approach achieves a significant improvement over baselines including the existing tree-structured CNNs with class-based grouping.

We present an efficient framework that can generate a coherent paragraph to describe a given video. Previous works on video captioning usually focus on video clips. They typically treat an entire video as a whole and generate the caption conditioned on a single embedding. On the contrary, we consider videos with rich temporal structures and aim to generate paragraph descriptions that can preserve the story flow while being coherent and concise. Towards this goal, we propose a new approach, which produces a descriptive paragraph by assembling temporally localized descriptions. Given a video, it selects a sequence of distinctive clips and generates sentences thereon in a coherent manner. Particularly, the selection of clips and the production of sentences are done jointly and progressively driven by a recurrent network – what to describe next depends on what have been said before. Here, the recurrent network is learned via self-critical sequence training with both sentence-level and paragraph-level rewards. On the ActivityNet Captions dataset, our method demonstrated the capability of generating high-quality paragraph descriptions for videos. Compared to those by other methods, the descriptions produced by our method are often more relevant, more coherent, and more concise.

Monocular depth estimation aims at estimating a pixelwise depth map for a single image, which has wide applications in scene understanding and autonomous driving. Existing supervised and unsupervised methods face great challenges. Supervised methods require large amounts of depth measurement data, which are generally difficult to obtain, while unsupervised methods are usually limited in estimation accuracy. Synthetic data generated by graphics engines provide a possible solution for collecting large amounts of depth data. However, the large domain gaps between synthetic and realistic data make directly training with them challenging. In this paper, we propose to use the stereo matching network as a proxy to learn depth from synthetic data and use predicted stereo disparity maps for supervising the monocular depth estimation network. Cross-domain synthetic data could be fully utilized in this novel framework. Different strategies are proposed to ensure learned depth perception capability well transferred across different domains. Our extensive experiments show state-of-the-art results of monocular depth estimation on KITTI dataset.

We address the problem of video object segmentation which outputs the masks of a target object throughout a video given only a bounding box in the first frame. There are two main challenges to this task. First, the background may contain similar objects as the target. Second, the appearance of the target object may change drastically over time. To tackle these challenges, we propose an end-to-end training network which accomplishes foreground predictions by leveraging the location-sensitive embeddings which are capable to distinguish the pixels of similar objects. To deal with appearance changes, for a test video, we propose a robust model adaptation method which pre-scans the whole video, generates pseudo foreground/background labels and retrains the model based on the labels. Our method outperforms the state-of-the-art methods on the DAVIS and the SegTrack v2 datasets.

Recent breakthroughs in Neural Architectural Search (NAS) have achieved state-of-the-art performances in applications such as image classification and language modeling. However, these techniques typically ignore device-related objectives such as inference time, memory usage, and power consumption. Optimizing neural architecture for device-related objectives is immensely crucial for deploying deep networks on portable devices with limited computing resources. We propose DPP-Net: Device-aware Progressive Search for Pareto-optimal Neural Architectures, optimizing for both device-related (e.g., inference time and memory usage) and device-agnostic (e.g., accuracy and model size) objectives. DPP-Net employs a compact search space inspired by current state-of-the-art mobile CNNs, and further improves search efficiency by adopting progressive search (Liu et al. 2017). Experimental results on CIFAR-10 are poised to demonstrate the effectiveness of Pareto-optimal networks found by DPP-Net, for three different devices: (1) a workstation with Titan X GPU, (2) NVIDIA Jetson TX1 embedded system, and (3) mobile phone with ARM Cortex-A53. Compared to CondenseNet and NASNet (Mobile), DPP-Net achieves better performances: higher accuracy & shorter inference time on various devices. Additional experimental results show that models found by DPP-Net also achieve considerably-good performance on ImageNet as well.

Incremental learning (il) has received a lot of attention recently, however, the literature lacks a precise problem definition, proper evaluation settings, and metrics tailored specifically for the il problem. One of the main objectives of this work is to fill these gaps so as to provide a common ground for better understanding of il. The main challenge for an il algorithm is to update the classifier whilst preserving existing knowledge. We observe that, in addition to forgetting, a known issue while preserving knowledge, il also suffers from a problem we call intransigence, its inability to update knowledge. We introduce two metrics to quantify forgetting and intransigence that allow us to understand, analyse, and gain better insights into the behaviour of il algorithms. Furthermore, we present RWalk, a generalization of ewc++ (our efficient version of ewc [6]) and Path Integral [25] with a theoretically grounded KL-divergence based perspective. We provide a thorough analysis of various il algorithms on MNIST and CIFAR-100 datasets. In these experiments, RWalk obtains superior results in terms of accuracy, and also provides a better trade-off for forgetting and intransigence.

This paper targets the problem of image set-based face verification and identification. Unlike traditional single media (an image or video) setting, we encounter a set of heterogeneous contents containing orderless images and videos. The importance of each image is usually considered either equal or based on their independent quality assessment. 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 the latent space. Specifically, we first present a dependency-aware attention control (DAC) network, which resorts to actor-critic reinforcement learning for sequential attention decision of each image embedding to fully exploit the rich correlation cues among the unordered images. Moreover, we introduce its sample-efficient variant with off-policy experience replay to speed up the learning process. The pose-guided representation scheme can further boost the performance at the extremes of the pose variation.

We present a convolutional autoencoder that enables high fidelity volumetric reconstructions of human performance to be captured from multi-view video comprising only a small set of camera views. Our method yields similar end-to-end reconstruction error to that of a probabilistic visual hull computed using significantly more (double or more) viewpoints. We use a deep prior implicitly learned by the autoencoder trained over a dataset of view-ablated multi-view video footage of a wide range of subjects and actions. This opens up the possibility of high-end volumetric performance capture in on-set and prosumer scenarios where time or cost prohibit a high witness camera count.

We present a challenging and realistic novel dataset for evaluating 6-DOF object tracking algorithms. Existing datasets show serious limitations—notably, unrealistic synthetic data, or real data with large fiducial markers—preventing the community from obtaining an accurate picture of the state-of-the-art. Using a data acquisition pipeline based on a commercial motion capture system for acquiring accurate ground truth poses of real objects with respect to a Kinect V2 camera, we build a dataset which contains a total of 297 calibrated sequences. They are acquired in three different scenarios to evaluate the performance of trackers: stability, robustness to occlusion and accuracy during challenging interactions between a person and the object. We conduct an extensive study of a deep 6-DOF tracking architecture and determine a set of optimal parameters. We enhance the architecture and the training methodology to train a 6-DOF tracker that can robustly generalize to objects never seen during training, and demonstrate favorable performance compared to previous approaches trained specifically on the objects to track.

Subsurface scattering plays a significant role in determining the appearance of real-world surfaces. A light ray penetrating into the subsurface is repeatedly scattered and absorbed by particles along its path before reemerging from the outer interface, which determines its spectral radiance. We introduce a novel imaging method that enables the decomposition of the appearance of a fronto-parallel real-world surface into images of light with bounded path lengths, i.e., transient subsurface light transport. Our key idea is to observe each surface point under a variable ring light: a circular illumination pattern of increasingly larger radius centered on it. We show that the path length of light captured in each of these observations is naturally lower-bounded by the ring light radius. By taking the difference of ring light images of incrementally larger radii, we compute transient images that encode light with bounded path lengths. Experimental results on synthetic and complex real-world surfaces demonstrate that the recovered transient images reveal the subsurface structure of general translucent inhomogeneous surfaces. We further show that their differences reveal the surface colors at different surface depths. The proposed method is the first to enable the unveiling of dense and continuous subsurface structures from steady-state external appearance using ordinary camera and illumination.

In this paper we present an efficient modeling framework for large scale urban scenes. Taking surface meshes derived from multi-view-stereo systems as input, our algorithm outputs simplified models with semantics at different levels of detail (LODs). Our key observation is that urban building is usually composed of planar roof tops connected with vertical walls. There are two major steps in our framework: segmentation and building modeling. The scene is first segmented into four classes with a Markov random field combining height and image features. In the following modeling step, various 2D line segments sketching the roof boundaries are detected and slice the plane into faces. Through assigning each face with a roof plane, the final model is constructed by extruding the faces to the corresponding planes. By combining geometric and appearance cues together, the proposed method is robust and fast compared to the state-of-the-art algorithms.

We address the problem of segmenting an object given a natural language expression that describes it. Current techniques tackle this task by either (i) directly or recursively merging linguistic and visual information in the channel dimension and then performing convolutions; or by (ii) mapping the expression to a space in which it can be thought of as a filter, whose response is directly related to the presence of the object at a given spatial coordinate in the image, so that a convolution can be applied to look for the object. We propose a novel method that integrates these two insights in order to fully exploit the recursive nature of language. Additionally, during the upsampling process, we take advantage of the intermediate information generated when downsampling the image, so that detailed segmentations can be obtained. We compare our method against the state-of-the-art approaches in four standard datasets, in which it surpasses all previous methods in six of eight of the splits for this task.

The problem of single-view 3D shape completion or reconstruction is challenging, because among the many possible shapes that explain an observation, most are implausible and do not correspond to natural objects. Recent research in the field has tackled this problem by exploiting the expressiveness of deep convolutional networks. In fact, there is another level of ambiguity that is often overlooked: among plausible shapes, there are still multiple shapes that fit the 2D image equally well; i.e., the ground truth shape is non-deterministic given a single-view input. Existing fully supervised approaches fail to address this issue, and often produce blurry mean shapes with smooth surfaces but no fine details. In this paper, we propose ShapeHD, pushing the limit of single-view shape completion and reconstruction by integrating deep generative models with adversarially learned shape priors. The learned priors serve as a regularizer, penalizing the model only if its output is unrealistic, not if it deviates from the ground truth. Our design thus overcomes both levels of ambiguity aforementioned. Experiments demonstrate that ShapeHD outperforms state of the art by a large margin in both shape completion and shape reconstruction on multiple real datasets.

When intelligent agents learn visuomotor behaviors from human demonstrations, they may benefit from knowing where the human is allocating visual attention, which can be inferred from their gaze. A wealth of information regarding intelligent decision making is conveyed by human gaze allocation; hence, exploiting such information has the potential to improve the agents’ performance. With this motivation, we propose the AGIL (Attention Guided Imitation Learning) framework. We collect high-quality human action and gaze data while playing Atari games in a carefully controlled experimental setting. Using these data, we first train a deep neural network that can predict human gaze positions and visual attention with high accuracy (the gaze network) and then train another network to predict human actions (the policy network). Incorporating the learned attention model from the gaze network into the policy network significantly improves the action prediction accuracy and task performance.

For many computer vision applications, such as image description and human identification, recognizing the visual attributes of humans is an essential yet challenging problem. Its challenges originate from its multi-label nature, the large underlying class imbalance and the lack of spatial annotations. Existing methods follow either a computer vision approach while failing to account for class imbalance, or explore machine learning solutions, which disregard the spatial and semantic relations that exist in the images. With that in mind, we propose an effective method that extracts and aggregates visual attention masks at different scales. We introduce a loss function to handle class imbalance both at class and at an instance level and further demonstrate that penalizing attention masks with high prediction variance accounts for the weak supervision of the attention mechanism. By identifying and addressing these challenges, we achieve state-of-the-art results with a simple attention mechanism in both PETA and WIDER-Attribute datasets without additional context or side information.

The recent years have witnessed significant growth in constructing robust generative models to capture informative distributions of natural data. However, it is difficult to fully exploit the distribution of complex data, like images and videos, due to the high dimensionality of ambient space. Sequentially, how to effectively guide the training of generative models is a crucial issue. In this paper, we present a subspace-based generative adversarial network (Sub-GAN) which simultaneously disentangles multiple latent subspaces and generates diverse samples correspondingly. Since the high-dimensional natural data usually lies on a union of low-dimensional subspaces which contain semantically extensive structure, Sub-GAN incorporates a novel clusterer that can interact with the generator and discriminator via subspace information. Unlike the traditional generative models, the proposed Sub-GAN can control the diversity of the generated samples via the multiplicity of the learned subspaces. Moreover, the Sub-GAN follows an unsupervised fashion to explore not only the visual classes but the latent continuous attributes. We demonstrate that our model can discover meaningful visual attributes which is hard to be annotated via strong supervision, e.g., the writing style of digits, thus avoid the mode collapse problem. Extensive experimental results show the competitive performance of the proposed method for both generating diverse images with satisfied quality and discovering discriminative latent subspaces.

This paper proposes a fast video salient object detection model, based on a novel recurrent network architecture, named Pyramid Dilated Bidirectional ConvLSTM (PDB-ConvLSTM). A Pyramid Dilated Convolution (PDC) module is first designed for simultaneously extracting spatial features at multiple scales. These spatial features are then concatenated and fed into an extended Deeper Bidirectional ConvLSTM (DB-ConvLSTM) to learn spatiotemporal information. Forward and backward ConvLSTM units are placed in two layers and connected in a cascaded way, encouraging information flow between the bi-directional streams and leading to deeper feature extraction. We further augment DB-ConvLSTM with a PDC-like structure, by adopting several dilated DB-ConvLSTMs to extract multi-scale spatiotemporal information. Extensive experimental results show that our method outperforms previous video saliency models in a large margin, with a real-time speed of 20 fps on a single GPU. With unsupervised video object segmentation as an example application, the proposed model (with a CRF-based post-process) achieves state-of-the-art results on two popular benchmarks, well demonstrating its superior performance and high applicability.

Simultaneously and accurately forecasting the behavior of many interacting agents is imperative for computer vision applications to be widely deployed (e.g., autonomous vehicles, security, surveillance, sports). In this paper, we present a technique using conditional variational autoencoder which learns a model that “personalizes” prediction to individual agent behavior within a group representation. Given the volume of data available and its adversarial nature, we focus on the sport of basketball and show that our approach efficiently predicts context-specific agent motions. We find that our model generates results that are three times as accurate as previous state of the art approaches (5.74 ft vs. 17.95 ft).

Existing deblurring methods mainly focus on developing effective image priors and assume that blurred images contain insignificant amounts of noise. However, state-of-the-art deblurring methods do not perform well on real-world images degraded with significant noise or outliers. To address these issues, we show that it is critical to learn data fitting terms beyond the commonly used $$\ell _1$$ ℓ 1 or $$\ell _2$$ ℓ 2 norm. We propose a simple and effective discriminative framework to learn data terms that can adaptively handle blurred images in the presence of severe noise and outliers. Instead of learning the distribution of the data fitting errors, we directly learn the associated shrinkage function for the data term using a cascaded architecture, which is more flexible and efficient. Our analysis shows that the shrinkage functions learned at the intermediate stages can effectively suppress noise and preserve image structures. Extensive experimental results show that the proposed algorithm performs favorably against state-of-the-art methods.

Domain adaptation is an important tool to transfer knowledge about a task (e.g. classification) learned in a source domain to a second, or target domain. Current approaches assume that task-relevant target-domain data is available during training. We demonstrate how to perform domain adaptation when no such task-relevant target-domain data is available. To tackle this issue, we propose zero-shot deep domain adaptation (ZDDA), which uses privileged information from task-irrelevant dual-domain pairs. ZDDA learns a source-domain representation which is not only tailored for the task of interest but also close to the target-domain representation. Therefore, the source-domain task of interest solution (e.g. a classifier for classification tasks) which is jointly trained with the source-domain representation can be applicable to both the source and target representations. Using the MNIST, Fashion-MNIST, NIST, EMNIST, and SUN RGB-D datasets, we show that ZDDA can perform domain adaptation in classification tasks without access to task-relevant target-domain training data. We also extend ZDDA to perform sensor fusion in the SUN RGB-D scene classification task by simulating task-relevant target-domain representations with task-relevant source-domain data. To the best of our knowledge, ZDDA is the first domain adaptation and sensor fusion method which requires no task-relevant target-domain data. The underlying principle is not particular to computer vision data, but should be extensible to other domains.

The objective of this work is set-based verification, e.g. to decide if two sets of images of a face are of the same person or not. The traditional approach to this problem is to learn to generate a feature vector per image, aggregate them into one vector to represent the set, and then compute the cosine similarity between sets. Instead, we design a neural network architecture that can directly learn set-wise verification.Our contributions are: (i) We propose a Deep Comparator Network (DCN) that can ingest a pair of sets (each may contain a variable number of images) as inputs, and compute a similarity between the pair – this involves attending to multiple discriminative local regions (landmarks), and comparing local descriptors between pairs of faces; (ii) To encourage high-quality representations for each set, internal competition is introduced for recalibration based on the landmark score; (iii) Inspired by image retrieval, a novel hard sample mining regime is proposed to control the sampling process, such that the DCN is complementary to the standard image classification models. Evaluations on the IARPA Janus face recognition benchmarks show that the comparator networks outperform the previous state-of-the-art results by a large margin.

In this paper, we propose a novel object detection framework named “Deep Regionlets” by establishing a bridge between deep neural networks and conventional detection schema for accurate generic object detection. Motivated by the abilities of regionlets for modeling object deformation and multiple aspect ratios, we incorporate regionlets into an end-to-end trainable deep learning framework. The deep regionlets framework consists of a region selection network and a deep regionlet learning module. Specifically, given a detection bounding box proposal, the region selection network provides guidance on where to select regions to learn the features from. The regionlet learning module focuses on local feature selection and transformation to alleviate local variations. To this end, we first realize non-rectangular region selection within the detection framework to accommodate variations in object appearance. Moreover, we design a “gating network” within the regionlet leaning module to enable soft regionlet selection and pooling. The Deep Regionlets framework is trained end-to-end without additional efforts. We perform ablation studies and conduct extensive experiments on the PASCAL VOC and Microsoft COCO datasets. The proposed framework outperforms state-of-the-art algorithms, such as RetinaNet and Mask R-CNN, even without additional segmentation labels.