CVPR 2017 Papers

Supervised deep learning often suffers from the lack of sufficient training data. Specifically in the context of monocular depth map prediction, it is barely possible to determine dense ground truth depth images in realistic dynamic outdoor environments. When using LiDAR sensors, for instance, noise is present in the distance measurements, the calibration between sensors cannot be perfect, and the measurements are typically much sparser than the camera images. In this paper, we propose a novel approach to depth map prediction from monocular images that learns in a semi-supervised way. While we use sparse ground-truth depth for supervised learning, we also enforce our deep network to produce photoconsistent dense depth maps in a stereo setup using a direct image alignment loss. In experiments we demonstrate superior performance in depth map prediction from single images compared to the state-of-the-art methods.

3D reconstruction from shading information through Photometric Stereo is considered a very challenging problem in Computer Vision. Although this technique can potentially provide highly detailed shape recovery, its accuracy is critically dependent on a numerous set of factors among them the reliability of the light sources in emitting a constant amount of light. In this work, we propose a novel variational approach to solve the so called semi-calibrated near field Photometric Stereo problem, where the positions but not the brightness of the light sources are known. Additionally, we take into account realistic modeling features such as perspective viewing geometry and heterogeneous scene composition, containing both diffuse and specular objects. Furthermore, we also relax the point light source assumption that usually constraints the near field formulation by explicitly calculating the light attenuation maps. Synthetic experiments are performed for quantitative evaluation for a wide range of cases whilst real experiments provide comparisons, qualitatively outperforming the state of the art.

The role of semantics in zero-shot learning is considered. The effectiveness of previous approaches is analyzed according to the form of supervision provided. While some learn semantics independently, others only supervise the semantic subspace explained by training classes. Thus, the former is able to constrain the whole space but lacks the ability to model semantic correlations. The latter addresses this issue but leaves part of the semantic space unsupervised. This complementarity is exploited in a new convolutional neural network (CNN) framework, which proposes the use of semantics as constraints for recognition.Although a CNN trained for classification has no transfer ability, this can be encouraged by learning an hidden semantic layer together with a semantic code for classification. Two forms of semantic constraints are then introduced. The first is a loss-based regularizer that introduces a generalization constraint on each semantic predictor. The second is a codeword regularizer that favors semantic-to-class mappings consistent with prior semantic knowledge while allowing these to be learned from data. Significant improvements over the state-of-the-art are achieved on several datasets.

In this paper we propose a framework for spatially and temporally coherent semantic co-segmentation and reconstruction of complex dynamic scenes from multiple static or moving cameras. Semantic co-segmentation exploits the coherence in semantic class labels both spatially, between views at a single time instant, and temporally, between widely spaced time instants of dynamic objects with similar shape and appearance. We demonstrate that semantic coherence results in improved segmentation and reconstruction for complex scenes. A joint formulation is proposed for semantically coherent object-based co-segmentation and reconstruction of scenes by enforcing consistent semantic labelling between views and over time. Semantic tracklets are introduced to enforce temporal coherence in semantic labelling and reconstruction between widely spaced instances of dynamic objects. Tracklets of dynamic objects enable unsupervised learning of appearance and shape priors that are exploited in joint segmentation and reconstruction. Evaluation on challenging indoor and outdoor sequences with hand-held moving cameras shows improved accuracy in segmentation, temporally coherent semantic labelling and 3D reconstruction of dynamic scenes.

This paper describes a fast and accurate semantic image segmentation approach that encodes not only segmentation-specified features but also high-order context compatibilities and boundary guidance constraints. We introduce a structured patch prediction technique to make a trade-off between classification discriminability and boundary sensibility for features. Both label and feature contexts are embedded to ensure recognition accuracy and compatibility, while the complexity of the high order cliques is reduced by a distance-aware sampling and pooling strategy. The proposed joint model also employs a guidance CRF to further enhance the segmentation performance. The message passing step is augmented with the guided filtering which enables an efficient and joint training of the whole system in an end-to-end fashion. Our proposed joint model outperforms the state-of-art on Pascal VOC 2012 and Cityscapes, with mIoU(%) of 82.5 and 79.2 respectively. It also reaches a leading performance on ADE20K, which is the dataset of the scene parsing track in ILSVRC 2016.

This paper focuses on semantic scene completion, a task for producing a complete 3D voxel representation of volumetric occupancy and semantic labels for a scene from a single-view depth map observation. Previous work has considered scene completion and semantic labeling of depth maps separately. However, we observe that these two problems are tightly intertwined. To leverage the coupled nature of these two tasks, we introduce the semantic scene completion network (SSCNet), an end-to-end 3D convolutional network that takes a single depth image as input and simultaneously outputs occupancy and semantic labels for all voxels in the camera view frustum. Our network uses a dilation-based 3D context module to efficiently expand the receptive field and enable 3D context learning. To train our network, we construct SUNCG - a manually created largescale dataset of synthetic 3D scenes with dense volumetric annotations. Our experiments demonstrate that the joint model outperforms methods addressing each task in isolation and outperforms alternative approaches on the semantic scene completion task. The dataset and code is available at http://sscnet.cs.princeton.edu.

The "CNN-RNN" design pattern is increasingly widely applied in a variety of image annotation tasks including multi-label classification and captioning. Existing models use the weakly semantic CNN hidden layer or its transform as the image embedding that provides the interface between the CNN and RNN. This leaves the RNN overstretched with two jobs: predicting the visual concepts and modelling their correlations for generating structured annotation output. Importantly this makes the end-to-end training of the CNN and RNN slow and ineffective due to the difficulty of back propagating gradients through the RNN to train the CNN. We propose a simple modification to the design pattern that makes learning more effective and efficient. Specifically, we propose to use a semantically regularised embedding layer as the interface between the CNN and RNN. Regularising the interface can partially or completely decouple the learning problems, allowing each to be more effectively trained and jointly training much more efficient. Extensive experiments show that state-of-the art performance is achieved on multi-label classification as well as image captioning.

Dense 3D reconstruction from RGB images is a highly ill-posed problem due to occlusions, textureless or reflective surfaces, as well as other challenges. We propose object-level shape priors to address these ambiguities. Towards this goal, we formulate a probabilistic model that integrates multi-view image evidence with 3D shape information from multiple objects. Inference in this model yields a dense 3D reconstruction of the scene as well as the existence and precise 3D pose of the objects in it. Our approach is able to recover fine details not captured in the input shapes while defaulting to the input models in occluded regions where image evidence is weak. Due to its probabilistic nature, the approach is able to cope with the approximate geometry of the 3D models as well as input shapes that are not present in the scene. We evaluate the approach quantitatively on several challenging indoor and outdoor datasets.

Semantic image inpainting is a challenging task where large missing regions have to be filled based on the available visual data. Existing methods which extract information from only a single image generally produce unsatisfactory results due to the lack of high level context. In this paper, we propose a novel method for semantic image inpainting, which generates the missing content by conditioning on the available data. Given a trained generative model, we search for the closest encoding of the corrupted image in the latent image manifold using our context and prior losses. This encoding is then passed through the generative model to infer the missing content. In our method, inference is possible irrespective of how the missing content is structured, while the state-of-the-art learning based method requires specific information about the holes in the training phase. Experiments on three datasets show that our method successfully predicts information in large missing regions and achieves pixel-level photorealism, significantly outperforming the state-of-the-art methods.

A Semantic Compositional Network (SCN) is developed for image captioning, in which semantic concepts (i.e., tags) are detected from the image, and the probability of each tag is used to compose the parameters in a long short-term memory (LSTM) network. The SCN extends each weight matrix of the LSTM to an ensemble of tag-dependent weight matrices. The degree to which each member of the ensemble is used to generate an image caption is tied to the image-dependent probability of the corresponding tag. In addition to captioning images, we also extend the SCN to generate captions for video clips. We qualitatively analyze semantic composition in SCNs, and quantitatively evaluate the algorithm on three benchmark datasets: COCO, Flickr30k, and Youtube2Text. Experimental results show that the proposed method significantly outperforms prior state-of-the-art approaches, across multiple evaluation metrics.

Existing zero-shot learning (ZSL) models typically learn a projection function from a feature space to a semantic embedding space (e.g. attribute space). However, such a projection function is only concerned with predicting the training seen class semantic representation (e.g. attribute prediction) or classification. When applied to test data, which in the context of ZSL contains different (unseen) classes without training data, a ZSL model typically suffers from the project domain shift problem. In this work, we present a novel solution to ZSL based on learning a Semantic AutoEncoder (SAE). Taking the encoder-decoder paradigm, an encoder aims to project a visual feature vector into the semantic space as in the existing ZSL models. However, the decoder exerts an additional constraint, that is, the projection/code must be able to reconstruct the original visual feature. We show that with this additional reconstruction constraint, the learned projection function from the seen classes is able to generalise better to the new unseen classes. Importantly, the encoder and decoder are linear and symmetric which enable us to develop an extremely efficient learning algorithm. Extensive experiments on six benchmark datasets demonstrate that the proposed SAE outperforms significantly the existing ZSL models with the additional benefit of lower computational cost. Furthermore, when the SAE is applied to supervised clustering problem, it also beats the state-of-the-art.

Common visual recognition tasks such as classification, object detection, and semantic segmentation are rapidly reaching maturity, and given the recent rate of progress, it is not unreasonable to conjecture that techniques for many of these problems will approach human levels of performance in the next few years. In this paper we look to the future: what is the next frontier in visual recognition? We offer one possible answer to this question. We propose a detailed image annotation that captures information beyond the visible pixels and requires complex reasoning about full scene structure. Specifically, we create an amodal segmentation of each image: the full extent of each region is marked, not just the visible pixels. Annotators outline and name all salient regions in the image and specify a partial depth order. The result is a rich scene structure, including visible and occluded portions of each region, figure-ground edge information, semantic labels, and object overlap. We create two datasets for semantic amodal segmentation. First, we label 500 images in the BSDS dataset with multiple annotators per image, allowing us to study the statistics of human annotations. We show that the proposed full scene annotation is surprisingly consistent between annotators, including for regions and edges. Second, we annotate 5000 images from COCO. This larger dataset allows us to explore a number of algorithmic ideas for amodal segmentation and depth ordering. We introduce novel metrics for these tasks, and along with our strong baselines, define concrete new challenges for the community.

We propose a new self-supervised CNN pre-training technique based on a novel auxiliary task called odd-one-out learning. In this task, the machine is asked to identify the unrelated or odd element from a set of otherwise related elements. We apply this technique to self-supervised video representation learning where we sample subsequences from videos and ask the network to learn to predict the odd video subsequence. The odd video subsequence is sampled such that it has wrong temporal order of frames while the even ones have the correct temporal order. Therefore, to generate a odd-one-out question no manual annotation is required. Our learning machine is implemented as multi-stream convolutional neural network, which is learned end-to-end. Using odd-one-out networks, we learn temporal representations for videos that generalizes to other related tasks such as action recognition. On action classification, our method obtains 60.3% on the UCF101 dataset using only UCF101 data for training which is approximately 10% better than current state-of-the-art self-supervised learning methods. Similarly, on HMDB51 dataset we outperform self-supervised state-of-the art methods by 12.7% on action classification task.

End-to-end training from scratch of current deep architectures for new computer vision problems would require Imagenet-scale datasets, and this is not always possible. In this paper we present a method that is able to take advantage of freely available multi-modal content to train computer vision algorithms without human supervision. We put forward the idea of performing self-supervised learning of visual features by mining a large scale corpus of multi-modal (text and image) documents. We show that discriminative visual features can be learnt efficiently by training a CNN to predict the semantic context in which a particular image is more probable to appear as an illustration. For this we leverage the hidden semantic structures discovered in the text corpus with a well-known topic modeling technique. Our experiments demonstrate state of the art performance in image classification, object detection, and multi-modal retrieval compared to recent self-supervised or natural-supervised approaches.

In this paper, a self-learning approach is proposed towards solving scene-specific pedestrian detection problem without any human' annotation involved. The self-learning approach is deployed as progressive steps of object discovery, object enforcement, and label propagation. In the learning procedure, object locations in each frame are treated as latent variables that are solved with a progressive latent model (PLM). Compared with conventional latent models, the proposed PLM incorporates a spatial regularization term to reduce ambiguities in object proposals and to enforce object localization, and also a graph-based label propagation to discover harder instances in adjacent frames. With the difference of convex (DC) objective functions, PLM can be efficiently optimized with a concave-convex programming and thus guaranteeing the stability of self-learning. Extensive experiments demonstrate that even without annotation the proposed self-learning approach outperforms weakly supervised learning approaches, while achieving comparable performance with transfer learning and fully supervised approaches.

Recently it has been shown that policy-gradient methods for reinforcement learning can be utilized to train deep end-to-end systems directly on non-differentiable metrics for the task at hand. In this paper we consider the problem of optimizing image captioning systems using reinforcement learning, and show that by carefully optimizing our systems using the test metrics of the MSCOCO task, significant gains in performance can be realized. Our systems are built using a new optimization approach that we call self-critical sequence training (SCST). SCST is a form of the popular REINFORCE algorithm that, rather than estimating a "baseline" to normalize the rewards and reduce variance, utilizes the output of its own test-time inference algorithm to normalize the rewards it experiences. Using this approach, estimating the reward signal (as actor-critic methods must do) and estimating normalization (as REINFORCE algorithms typically do) is avoided, while at the same time harmonizing the model with respect to its test-time inference procedure. Empirically we find that directly optimizing the CIDEr metric with SCST and greedy decoding at test-time is highly effective. Our results on the MSCOCO evaluation sever establish a new state-of-the-art on the task, improving the best result in terms of CIDEr from 104.9 to 114.7.

In this paper we consider critical motion sequences (CMSs) of rolling-shutter (RS) SfM. Employing an RS camera model with linearized pure rotation, we show that the RS distortion can be approximately expressed by two internal parameters of an "imaginary" camera plus one-parameter nonlinear transformation similar to lens distortion. We then reformulate the problem as self-calibration of the imaginary camera, in which its skew and aspect ratio are unknown and varying in the image sequence. In the formulation, we derive a general representation of CMSs. We also show that our method can explain the CMS that was recently reported in the literature, and then present a new remedy to deal with the degeneracy. Our theoretical results agree well with experimental results; it explains degeneracies observed when we employ naive bundle adjustment, and how they are resolved by our method.

Most of computer vision focuses on what is in an image. We propose to train a standalone object-centric context representation to perform the opposite task: seeing what is not there. Given an image, our context model can predict where objects should exist, even when no object instances are present. Combined with object detection results, we can perform a novel vision task: finding where objects are missing in an image. Our model is based on a convolutional neural network structure. With a specially designed training strategy, the model learns to ignore objects and focus on context only. It is fully convolutional thus highly efficient. Experiments show the effectiveness of the proposed approach in one important accessibility task: finding city street regions where curb ramps are missing, which could help millions of people with mobility disabilities.

Understanding the camera wearer's activity is central to egocentric vision, yet one key facet of that activity is inherently invisible to the camera--the wearer's body pose. Prior work focuses on estimating the pose of hands and arms when they come into view, but this 1) gives an incomplete view of the full body posture, and 2) prevents any pose estimate at all in many frames, since the hands are only visible in a fraction of daily life activities. We propose to infer the "invisible pose" of a person behind the egocentric camera. Given a single video, our efficient learning-based approach returns the full body 3D joint positions for each frame. Our method exploits cues from the dynamic motion signatures of the surrounding scene--which change predictably as a function of body pose--as well as static scene structures that reveal the viewpoint (e.g., sitting vs. standing). We further introduce a novel energy minimization scheme to infer the pose sequence. It uses soft predictions of the poses per time instant together with a non-parametric model of human pose dynamics over longer windows. Our method outperforms an array of possible alternatives, including typical deep learning approaches for direct pose regression from images.