Generative adversarial networks (GANs) are a learning framework that rely on training a discriminator to estimate a measure of difference between a target and generated distributions. GANs, as normally formulated, rely on the generated samples being completely differentiable w.r.t. the generative parameters, and thus do not work for discrete data. We introduce a method for training GANs with discrete data that uses the estimated difference measure from the discriminator to compute importance weights for generated samples, thus providing a policy gradient for training the generator. The importance weights have a strong connection to the decision boundary of the discriminator, and we call our method boundary-seeking GANs (BGANs). We demonstrate the effectiveness of the proposed algorithm with discrete image and character-based natural language generation. In addition, the boundary-seeking objective extends to continuous data, which can be used to improve stability of training, and we demonstrate this on Celeba, Large-scale Scene Understanding (LSUN) bedrooms, and Imagenet without conditioning.
This paper addresses a fundamental problem of scene understanding: How to parse the scene image into a structured configuration (i.e., a semantic object hierarchy with object interaction relations) that finely accords with human perception. We propose a deep architecture consisting of two networks: i) a convolutional neural network (CNN) extracting the image representation for pixelwise object labeling and ii) a recursive neural network (RNN) discovering the hierarchical object structure and the inter-object relations. Rather than relying on elaborative user annotations (e.g., manually labeling semantic maps and relations), we train our deep model in a weakly-supervised manner by leveraging the descriptive sentences of the training images. Specifically, we decompose each sentence into a semantic tree consisting of nouns and verb phrases, and facilitate these trees discovering the configurations of the training images. Once these scene configurations are determined, then the parameters of both the CNN and RNN are updated accordingly by back propagation. The entire model training is accomplished through an Expectation-Maximization method. Extensive experiments suggest that our model is capable of producing meaningful and structured scene configurations and achieving more favorable scene labeling performance on PASCAL VOC 2012 over other state-of-the-art weakly-supervised methods.
Deep learning models with convolutional and recurrent networks are now ubiquitous and analyze massive amounts of audio, image, video, text and graph data, with applications in automatic translation, speech-to-text, scene understanding, ranking user preferences, ad placement, etc. Competing frameworks for building these networks such as TensorFlow, Chainer, CNTK, Torch/PyTorch, Caffe1/2, MXNet and Theano, explore different tradeoffs between usability and expressiveness, research or production orientation and supported hardware. They operate on a DAG of computational operators, wrapping high-performance libraries such as CUDNN (for NVIDIA GPUs) or NNPACK (for various CPUs), and automate memory allocation, synchronization, distribution. Custom operators are needed where the computation does not fit existing high-performance library calls, usually at a high engineering cost. This is frequently required when new operators are invented by researchers: such operators suffer a severe performance penalty, which limits the pace of innovation. Furthermore, even if there is an existing runtime call these frameworks can use, it often doesn't offer optimal performance for a user's particular network architecture and dataset, missing optimizations between operators as well as optimizations that can be done knowing the size and shape of data. Our contributions include (1) a language close to the mathematics of deep learning called Tensor Comprehensions, (2) a polyhedral Just-In-Time compiler to convert a mathematical description of a deep learning DAG into a CUDA kernel with delegated memory management and synchronization, also providing optimizations such as operator fusion and specialization for specific sizes, (3) a compilation cache populated by an autotuner. [Abstract cutoff]
With the availability of low-cost and compact 2.5/3D visual sensing devices, computer vision community is experiencing a growing interest in visual scene understanding of indoor environments. This survey paper provides a comprehensive background to this research topic. We begin with a historical perspective, followed by popular 3D data representations and a comparative analysis of available datasets. Before delving into the application specific details, this survey provides a succinct introduction to the core technologies that are the underlying methods extensively used in the literature. Afterwards, we review the developed techniques according to a taxonomy based on the scene understanding tasks. This covers holistic indoor scene understanding as well as subtasks such as scene classification, object detection, pose estimation, semantic segmentation, 3D reconstruction, saliency detection, physics-based reasoning and affordance prediction. Later on, we summarize the performance metrics used for evaluation in different tasks and a quantitative comparison among the recent state-of-the-art techniques. We conclude this review with the current challenges and an outlook towards the open research problems requiring further investigation.
In Augmented Reality (AR) environment, realistic interactions between the virtual and real objects play a crucial role in user experience. Much of recent advances in AR has been largely focused on developing geometry-aware environment, but little has been done in dealing with interactions at the semantic level. High-level scene understanding and semantic descriptions in AR would allow effective design of complex applications and enhanced user experience. In this paper, we present a novel approach and a prototype system that enables the deeper understanding of semantic properties of the real world environment, so that realistic physical interactions between the real and the virtual objects can be generated. A material-aware AR environment has been created based on the deep material learning using a fully convolutional network (FCN). The state-of-the-art dense Simultaneous Localisation and Mapping (SLAM) has been used for the semantic mapping. Together with efficient accelerated 3D ray casting, natural and realistic physical interactions are generated for interactive AR games. Our approach has significant impact on the future development of advanced AR systems and applications.
Appropriate modeling of a surveillance scene is essential for detection of anomalies in road traffic. Learning usual paths can provide valuable insight into road traffic conditions and thus can help in identifying unusual routes taken by commuters/vehicles. If usual traffic paths are learned in a nonparametric way, manual interventions in road marking road can be avoided. In this paper, we propose an unsupervised and nonparametric method to learn frequently used paths from the tracks of moving objects in $\Theta(kn)$ time, where $k$ denotes the number of paths and $n$ represents the number of tracks. In the proposed method, temporal dependencies of the moving objects are considered to make the clustering meaningful using Temporally Incremental Gravity Model (TIGM). In addition, the distance-based scene learning makes it intuitive to estimate the model parameters. Further, we have extended TIGM hierarchically as Dynamically Evolving Model (DEM) to represent notable traffic dynamics of a scene. Experimental validation reveals that the proposed method can learn a scene quickly without prior knowledge about the number of paths ($k$). We have compared the results with various state-of-the-art methods. We have also highlighted the advantages of the proposed method over existing techniques popularly used for designing traffic monitoring applications. It can be used for administrative decision making to control traffic at junctions or crowded places and generate alarm signals, if necessary.
This paper proposes a segmentation-free, automatic and efficient procedure to detect general geometric quadric forms in point clouds, where clutter and occlusions are inevitable. Our everyday world is dominated by man-made objects which are designed using 3D primitives (such as planes, cones, spheres, cylinders, etc.). These objects are also omnipresent in industrial environments. This gives rise to the possibility of abstracting 3D scenes through primitives, thereby positions these geometric forms as an integral part of perception and high level 3D scene understanding. As opposed to state-of-the-art, where a tailored algorithm treats each primitive type separately, we propose to encapsulate all types in a single robust detection procedure. At the center of our approach lies a closed form 3D quadric fit, operating in both primal & dual spaces and requiring as low as 4 oriented-points. Around this fit, we design a novel, local null-space voting strategy to reduce the 4-point case to 3. Voting is coupled with the famous RANSAC and makes our algorithm orders of magnitude faster than its conventional counterparts. This is the first method capable of performing a generic cross-type multi-object primitive detection in difficult scenes. Results on synthetic and real datasets support the validity of our method.
Learning depth from a single image, as an important issue in scene understanding, has attracted a lot of attention in the past decade. The accuracy of the depth estimation has been improved from conditional Markov random fields, non-parametric methods, to deep convolutional neural networks most recently. However, there exist inherent ambiguities in recovering 3D from a single 2D image. In this paper, we first prove the ambiguity between the focal length and monocular depth learning, and verify the result using experiments, showing that the focal length has a great influence on accurate depth recovery. In order to learn monocular depth by embedding the focal length, we propose a method to generate synthetic varying-focal-length dataset from fixed-focal-length datasets, and a simple and effective method is implemented to fill the holes in the newly generated images. For the sake of accurate depth recovery, we propose a novel deep neural network to infer depth through effectively fusing the middle-level information on the fixed-focal-length dataset, which outperforms the state-of-the-art methods built on pre-trained VGG. Furthermore, the newly generated varying-focal-length dataset is taken as input to the proposed network in both learning and inference phases. Extensive experiments on the fixed- and varying-focal-length datasets demonstrate that the learned monocular depth with embedded focal length is significantly improved compared to that without embedding the focal length information.
As an agent moves through the world, the apparent motion of scene elements is (usually) inversely proportional to their depth. 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. This 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 those 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 (unsupervised) relative depth in the specific videos associated with various downstream tasks. 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.
Estimation of 3D motion in a dynamic scene from a temporal pair of images is a core task in many scene understanding problems. In real world applications, a dynamic scene is commonly captured by a moving camera (i.e., panning, tilting or hand-held), increasing the task complexity because the scene is observed from different view points. The main challenge is the disambiguation of the camera motion from scene motion, which becomes more difficult as the amount of rigidity observed decreases, even with successful estimation of 2D image correspondences. Compared to other state-of-the-art 3D scene flow estimation methods, in this paper we propose to \emph{learn} the rigidity of a scene in a supervised manner from a large collection of dynamic scene data, and directly infer a rigidity mask from two sequential images with depths. With the learned network, we show how we can effectively estimate camera motion and projected scene flow using computed 2D optical flow and the inferred rigidity mask. For training and testing the rigidity network, we also provide a new semi-synthetic dynamic scene dataset (synthetic foreground objects with a real background) and an evaluation split that accounts for the percentage of observed non-rigid pixels. Through our evaluation we show the proposed framework outperforms current state-of-the-art scene flow estimation methods in challenging dynamic scenes.
With the technological advancements of aerial imagery and accurate 3d reconstruction of urban environments, more and more attention has been paid to the automated analyses of urban areas. In our work, we examine two important aspects that allow live analysis of building structures in city models given oblique aerial imagery, namely automatic building extraction with convolutional neural networks (CNNs) and selective real-time depth estimation from aerial imagery. We use transfer learning to train the Faster R-CNN method for real-time deep object detection, by combining a large ground-based dataset for urban scene understanding with a smaller number of images from an aerial dataset. We achieve an average precision (AP) of about 80% for the task of building extraction on a selected evaluation dataset. Our evaluation focuses on both dataset-specific learning and transfer learning. Furthermore, we present an algorithm that allows for multi-view depth estimation from aerial imagery in real-time. We adopt the semi-global matching (SGM) optimization strategy to preserve sharp edges at object boundaries. In combination with the Faster R-CNN, it allows a selective reconstruction of buildings, identified with regions of interest (RoIs), from oblique aerial imagery.
In this paper, we demonstrate an end-to-end spatiotemporal gesture learning approach for 3D point cloud data using a new gestures dataset of point clouds acquired from a 3D sensor. Nine classes of gestures were learned from gestures sample data. We mapped point cloud data into dense occupancy grids, then time steps of the occupancy grids are used as inputs into a 3D convolutional neural network which learns the spatiotemporal features in the data without explicit modeling of gesture dynamics. We also introduced a 3D region of interest jittering approach for point cloud data augmentation. This resulted in an increased classification accuracy of up to 10% when the augmented data is added to the original training data. The developed model is able to classify gestures from the dataset with 84.44% accuracy. We propose that point cloud data will be a more viable data type for scene understanding and motion recognition, as 3D sensors become ubiquitous in years to come.
Automatically scene understanding using machine learning algorithms has been widely applied to different industries to reduce the cost of manual labor. Nowadays, insurance companies launch express vehicle insurance claim and settlement by allowing customers uploading pictures taken by mobile devices. This kind of insurance claim is treated as small claim and can be processed either manually or automatically in a quick fashion. However, due to the increasing amount of claims every day, system or people are likely to be fooled by repeated claims for identical case leading to big lost to insurance companies.Thus, an anti-fraud checking before processing the claim is necessary. We create the first data set of car damage images collected from internet and local parking lots. In addition, we proposed an approach to generate robust deep features by locating the damages accurately and efficiently in the images. The state-of-the-art real-time object detector YOLO \cite{redmon2016you}is modified to train and discover damage region as an important part of the pipeline. Both local and global deep features are extracted using VGG model\cite{Simonyan14c}, which are fused later for more robust system performance. Experiments show our approach is effective in preventing fraud claims as well as meet the requirement to speed up the insurance claim prepossessing.
State-of-the-art semantic segmentation approaches increase the receptive field of their models by using either a downsampling path composed of poolings/strided convolutions or successive dilated convolutions. However, it is not clear which operation leads to best results. In this paper, we systematically study the differences introduced by distinct receptive field enlargement methods and their impact on the performance of a novel architecture, called Fully Convolutional DenseResNet (FC-DRN). FC-DRN has a densely connected backbone composed of residual networks. Following standard image segmentation architectures, receptive field enlargement operations that change the representation level are interleaved among residual networks. This allows the model to exploit the benefits of both residual and dense connectivity patterns, namely: gradient flow, iterative refinement of representations, multi-scale feature combination and deep supervision. In order to highlight the potential of our model, we test it on the challenging CamVid urban scene understanding benchmark and make the following observations: 1) downsampling operations outperform dilations when the model is trained from scratch, 2) dilations are useful during the finetuning step of the model, 3) coarser representations require less refinement steps, and 4) ResNets (by model construction) are good regularizers, since they can reduce the model capacity when needed. Finally, we compare our architecture to alternative methods and report state-of-the-art result on the Camvid dataset, with at least twice fewer parameters.
Saliency prediction can benefit from training that involves scene understanding that may be tangential to the central task; this may include understanding places, spatial layout, objects or involve different datasets and their bias. One can combine models, but to do this in a sophisticated manner can be complex, and also result in unwieldy networks or produce competing objectives that are hard to balance. In this paper, we propose a scalable system to leverage multiple powerful deep CNN models to better extract visual features for saliency prediction. Our design differs from previous studies in that the whole system is trained in an almost end-to-end piece-wise fashion. The encoder and decoder components are separately trained to deal with complexity tied to the computational paradigm and required space. Furthermore, the encoder can contain more than one CNN model to extract features, and models can have different architectures or be pre-trained on different datasets. This parallel design yields a better computational paradigm overcoming limits to the variety of information or inference that can be combined at the encoder stage towards deeper networks and a more powerful encoding. Our network can be easily expanded almost without any additional cost, and other pre-trained CNN models can be incorporated availing a wider range of visual knowledge. We denote our expandable multi-layer network as EML-NET and our method achieves the state-of-the-art results on the public saliency benchmarks, SALICON, MIT300 and CAT2000.
In remote sensing, each sensor can provide complementary or reinforcing information. It is valuable to fuse outputs from multiple sensors to boost overall performance. Previous supervised fusion methods often require accurate labels for each pixel in the training data. However, in many remote sensing applications, pixel-level labels are difficult or infeasible to obtain. In addition, outputs from multiple sensors often have different resolution or modalities. For example, rasterized hyperspectral imagery presents data in a pixel grid while airborne Light Detection and Ranging (LiDAR) generates dense three-dimensional (3D) point clouds. It is often difficult to directly fuse such multi-modal, multi-resolution data. To address these challenges, we present a novel Multiple Instance Multi-Resolution Fusion (MIMRF) framework that can fuse multi-resolution and multi-modal sensor outputs while learning from automatically-generated, imprecisely-labeled data. Experiments were conducted on the MUUFL Gulfport hyperspectral and LiDAR data set and a remotely-sensed soybean and weed data set. Results show improved, consistent performance on scene understanding and agricultural applications when compared to traditional fusion methods.
Semantic understanding and localization are fundamental enablers of robot autonomy that have for the most part been tackled as disjoint problems. While deep learning has enabled recent breakthroughs across a wide spectrum of scene understanding tasks, its applicability to state estimation tasks has been limited due to the direct formulation that renders it incapable of encoding scene-specific constrains. In this work, we propose the VLocNet++ architecture that employs a multitask learning approach to exploit the inter-task relationship between learning semantics, regressing 6-DoF global pose and odometry, for the mutual benefit of each of these tasks. Our network overcomes the aforementioned limitation by simultaneously embedding geometric and semantic knowledge of the world into the pose regression network. We propose a novel adaptive weighted fusion layer to aggregate motion-specific temporal information and to fuse semantic features into the localization stream based on region activations. Furthermore, we propose a self-supervised warping technique that uses the relative motion to warp intermediate network representations in the segmentation stream for learning consistent semantics. Finally, we introduce a first-of-a-kind urban outdoor localization dataset with pixel-level semantic labels and multiple loops for training deep networks. Extensive experiments on the challenging Microsoft 7-Scenes benchmark and our DeepLoc dataset demonstrate that our approach exceeds the state-of-the-art outperforming local feature-based methods while simultaneously performing multiple tasks and exhibiting substantial robustness in challenging scenarios.
Objects often occlude each other in scenes; Inferring their appearance beyond their visible parts plays an important role in scene understanding, depth estimation, object interaction and manipulation. In this paper, we study the challenging problem of completing the appearance of occluded objects. Doing so requires knowing which pixels to paint (segmenting the invisible parts of objects) and what color to paint them (generating the invisible parts). Our proposed novel solution, SeGAN, jointly optimizes for both segmentation and generation of the invisible parts of objects. Our experimental results show that: (a) SeGAN can learn to generate the appearance of the occluded parts of objects; (b) SeGAN outperforms state-of-the-art segmentation baselines for the invisible parts of objects; (c) trained on synthetic photo realistic images, SeGAN can reliably segment natural images; (d) by reasoning about occluder occludee relations, our method can infer depth layering.
Depth estimation and scene parsing are two particularly important tasks in visual scene understanding. In this paper we tackle the problem of simultaneous depth estimation and scene parsing in a joint CNN. The task can be typically treated as a deep multi-task learning problem [42]. Different from previous methods directly optimizing multiple tasks given the input training data, this paper proposes a novel multi-task guided prediction-and-distillation network (PAD-Net), which first predicts a set of intermediate auxiliary tasks ranging from low level to high level, and then the predictions from these intermediate auxiliary tasks are utilized as multi-modal input via our proposed multi-modal distillation modules for the final tasks. During the joint learning, the intermediate tasks not only act as supervision for learning more robust deep representations but also provide rich multi-modal information for improving the final tasks. Extensive experiments are conducted on two challenging datasets (i.e. NYUD-v2 and Cityscapes) for both the depth estimation and scene parsing tasks, demonstrating the effectiveness of the proposed approach.
Image data has a great potential of helping conventional visual inspections of civil engineering structures due to the ease of data acquisition and the advantages in capturing visual information. A variety of techniques have been proposed to detect damages, such as cracks and spalling on a close-up image of a single component (columns and road surfaces etc.). However, these techniques commonly suffer from severe false-positives especially when the image includes multiple components of different structures. To reduce the false-positives and extract reliable information about the structures' conditions, detection and localization of critical structural components are important first steps preceding the damage assessment. This study aims at recognizing bridge structural and non-structural components from images of urban scenes. During the bridge component recognition, every image pixel is classified into one of the five classes (non-bridge, columns, beams and slabs, other structural, other nonstructural) by multi-scale convolutional neural networks (multi-scale CNNs). To reduce false-positives and get consistent labels, the component classifications are integrated with scene understanding by an additional classifier with 10 higher-level scene classes (building, greenery, person, pavement, signs and poles, vehicles, bridges, water, sky, and others). The bridge component recognition integrated with the scene understanding is compared with the naive approach without scene classification in terms of accuracy, false-positives and consistencies to demonstrate the effectiveness of the integrated approach.