CVPR 2016 Papers

We presents a method for future localization: to predict plausible future trajectories of ego-motion in egocentric stereo images. Our paths avoid obstacles, move between objects, even turn around a corner into space behind objects. As a byproduct of the predicted trajectories, we discover the empty space occluded by foreground objects. One key innovation is the creation of an EgoRetinal map, akin to an illustrated tourist map, that `rearranges' pixels taking into accounts depth information, the ground plane, and body motion direction, so that it allows motion planning and perception of objects on one image space. We learn to plan trajectories directly on this EgoRetinal map using first person experience of walking around in a variety of scenes. In a testing phase, given an novel scene, we find multiple hypotheses of future trajectories from the learned experience. We refine them by minimizing a cost function that describes compatibility between the obstacles in the EgoRetinal map and trajectories. We quantitatively evaluate our method to show predictive validity and apply to various real world daily activities including walking, shopping, and social interactions.

We propose a method to recover the shape of a 3D room from a full-view indoor panorama. Our algorithm can automatically infer a 3D shape from a collection of partially oriented superpixel facets and line segments. The core part of the algorithm is a constraint graph, which includes lines and superpixels as vertices, and encodes their geometric relations as edges. A novel approach is proposed to perform 3D reconstruction based on the constraint graph by solving all the geometric constraints as constrained linear least-squares. The selected constraints used for reconstruction are identified using an occlusion detection method with a Markov random field. Experiments show that our method can recover room shapes that can not be addressed by previous approaches. Our method is also efficient, that is, the inference time for each panorama is less than 1 minute.

We introduce the concept of dynamic image, a novel compact representation of videos useful for video analysis especially when convolutional neural networks (CNNs) are used. The dynamic image is based on the rank pooling concept and is obtained through the parameters of a ranking machine that encodes the temporal evolution of the frames of the video. Dynamic images are obtained by directly applying rank pooling on the raw image pixels of a video producing a single RGB image per video. This idea is simple but powerful as it enables the use of existing CNN models directly on video data with fine-tuning. We present an efficient and effective approximate rank pooling operator, speeding it up orders of magnitude compared to rank pooling. Our new approximate rank pooling CNN layer allows us to generalize dynamic images to dynamic feature maps and we demonstrate the power of our new representations on standard benchmarks in action recognition achieving state-of-the-art performance.

Multi-person event recognition is a challenging task, often with many people active in the scene but only a small subset contributing to an actual event. In this paper, we propose a model which learns to detect events in such videos while automatically "attending" to the people responsible for the event. Our model does not use explicit annotations regarding who or where those people are during training and testing. In particular, we track people in videos and use a recurrent neural network (RNN) to represent the track features. We learn time-varying attention weights to combine these features at each time-instant. The attended features are then processed using another RNN for event detection/classification. Since most video datasets with multiple people are restricted to a small number of videos, we also collected a new basketball dataset comprising 257 basketball games with 14K event annotations corresponding to 11 event classes. Our model outperforms state-of-the-art methods for both event classification and detection on this new dataset. Additionally, we show that the attention mechanism is able to consistently localize the relevant players.

We introduce the dense captioning task, which requires a computer vision system to both localize and describe salient regions in images in natural language. The dense captioning task generalizes object detection when the descriptions consist of a single word, and Image Captioning when one predicted region covers the full image. To address the localization and description task jointly we propose a Fully Convolutional Localization Network (FCLN) architecture that processes an image with a single, efficient forward pass, requires no external regions proposals, and can be trained end-to-end with a single round of optimization. The architecture is composed of a Convolutional Network, a novel dense localization layer, and Recurrent Neural Network language model that generates the label sequences. We evaluate our network on the Visual Genome dataset, which comprises 94,000 images and 4,100,000 region-grounded captions. We observe both speed and accuracy improvements over baselines based on current state of the art approaches in both generation and retrieval settings.

We propose a deep learning approach for finding dense correspondences between 3D scans of people. Our method requires only partial geometric information in the form of two depth maps or partial reconstructed surfaces, works for humans in arbitrary poses and wearing any clothing, does not require the two people to be scanned from similar viewpoints, and runs in real time. We use a deep convolutional neural network to train a feature descriptor on depth map pixels, but crucially, rather than training the network to solve the shape correspondence problem directly, we train it to solve a body region classification problem, modified to increase the smoothness of the learned descriptors near region boundaries. This approach ensures that nearby points on the human body are nearby in feature space, and vice versa, rendering the feature descriptor suitable for computing dense correspondences between the scans. We validate our method on real and synthetic data for both clothed and unclothed humans, and show that our correspondences are more robust than is possible with state-of-the-art unsupervised methods, and more accurate to those found using methods that require full watertight 3D geometry.

We propose an image super-resolution method (SR) using a deeply-recursive convolutional network (DRCN). Our network has a very deep recursive layer (up to 16 recursions). Increasing recursion depth can improve performance without introducing new parameters for additional convolutions. Albeit advantages, learning a DRCN is very hard with a standard gradient descent method due to exploding/ vanishing gradients. To ease the difficulty of training, we propose two extensions: recursive supervision and skip-connection. Our method outperforms previous methods by a large margin.

This paper addresses the problem of structured scene parsing, i.e., parsing the input scene into a configuration including hierarchical semantic objects with their interaction relations. 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 annotations (e.g., manually labeled 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.

Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers---8x deeper than VGG nets but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.

This work presents a new approach to learning a frame-based classifier on weakly labelled sequence data by embedding a CNN within an iterative EM algorithm. This allows the CNN to be trained on a vast number of example images when only loose sequence level information is available for the source videos. Although we demonstrate this in the context of hand shape recognition, the approach has wider application to any video recognition task where frame level labelling is not available. The iterative EM algorithm leverages the discriminative ability of the CNN to iteratively refine the frame level annotation and subsequent training of the CNN. By embedding the classifier within an EM framework the CNN can easily be trained on 1 million hand images. We demonstrate that the final classifier generalises over both individuals and data sets. The algorithm is evaluated on over 3000 manually labelled hand shape images of 60 different classes which will be released to the community. Furthermore, we demonstrate its use in continuous sign language recognition on two publicly available large sign language data sets, where it outperforms the current state-of-the-art by a large margin. To our knowledge no previous work has explored expectation maximization without Gaussian mixture models to exploit weak sequence labels for sign language recognition.

While recent deep neural network models have achieved promising results on the image captioning task, they rely largely on the availability of corpora with paired image and sentence captions to describe objects in context. In this work, we propose the Deep Compositional Captioner (DCC) to address the task of generating descriptions of novel objects which are not present in paired image-sentence datasets. Our method achieves this by leveraging large object recognition datasets and external text corpora and by transferring knowledge between semantically similar concepts. Current deep caption models can only describe objects contained in paired image-sentence corpora, despite the fact that they are pre-trained with large object recognition datasets, namely ImageNet. In contrast, our model can compose sentences that describe novel objects and their interactions with other objects. We demonstrate our model's ability to describe novel concepts by empirically evaluating its performance on MSCOCO and show qualitative results on ImageNet images of objects for which no paired image-caption data exist. Further, we extend our approach to generate descriptions of objects in video clips. Our results show that DCC has distinct advantages over existing image and video captioning approaches for generating descriptions of new objects in context.

We describe a method to automatically detect contours, i.e. lines along which the surface orientation sharply changes, in large-scale outdoor point clouds. Contours are important intermediate features for structuring point clouds and converting them into high-quality surface or solid models, and are extensively used in graphics and mapping applications. Yet, detecting them in unstructured, inhomogeneous point clouds turns out to be surprisingly difficult, and existing line detection algorithms largely fail. We approach contour extraction as a two-stage discriminative learning problem. In the first stage, a contour score for each individual point is predicted with a binary classifier, using a set of features extracted from the point's neighborhood. The contour scores serve as a basis to construct an overcomplete graph of candidate contours. The second stage selects an optimal set of contours from the candidates. This amounts to a further binary classification in a higher-order MRF, whose cliques encode a preference for connected contours and penalize loose ends. The method can handle point clouds >10^7 points in a couple of minutes, and vastly outperforms a baseline that performs Canny-style edge detection on a range image representation of the point cloud.

Recently, there have been many progresses for the problem of non-rigid structure reconstruction based on 2D trajectories, but it is still challenging to deal with complex deformations or restricted view ranges. Promising alternatives are the piecewise reconstruction approaches, which divide trajectories into several local parts and stitch their individual reconstructions to produce an entire 3D structure. These methods show the state-of-the-art performance, however, most of them are specialized for relatively smooth surfaces and some are quite complicated. Meanwhile, it has been reported numerously in the field of pattern recognition that obtaining consensus from many weak hypotheses can give a strong, powerful result. Inspired by these reports, in this paper, we push the concept of part-based reconstruction to the limit: Instead of considering the parts as explicitly-divided local patches, we draw a large number of small random trajectory sets. From their individual reconstructions, we pull out a statistic of each 3D point to retrieve a strong reconstruction, of which the procedure can be expressed as a sparse l_1-norm minimization problem. In order to resolve the reflection ambiguity between weak (and possibly bad) reconstructions, we propose a novel optimization framework which only involves a single eigenvalue decomposition. The proposed method can be applied to any type of data and outperforms the existing methods for the benchmark sequences, even though it is composed of a few, simple steps. Furthermore, it is easily parallelizable, which is another advantage.

Very long baseline interferometry (VLBI) is a technique for imaging celestial radio emissions by simultaneously observing a source from telescopes distributed across Earth. The challenges in reconstructing images from fine angular resolution VLBI data are immense. The data is extremely sparse and noisy, thus requiring statistical image models such as those designed in the computer vision community. In this paper we present a novel Bayesian approach for VLBI image reconstruction. While other methods often require careful tuning and parameter selection for different types of data, our method (CHIRP) produces good results under different settings such as low SNR or extended emission. The success of our method is demonstrated on realistic synthetic experiments as well as publicly available real data. We present this problem in a way that is accessible to members of the community, and provide a dataset website (vlbiimaging.csail.mit.edu) that facilitates controlled comparisons across algorithms.

Clustering is the task of grouping a set of objects so that objects in the same cluster are more similar to each other than to those in other clusters. The crucial step in most clustering algorithms is to find an appropriate similarity metric, which is both challenging and problem-dependent. Supervised clustering approaches, which can exploit labeled clustered training data that share a common metric with the test set, have thus been proposed. Unfortunately, current metric learning approaches for supervised clustering do not scale to large or even medium-sized datasets. In this paper, we propose a new structured Mahalanobis Distance Metric Learning method for supervised clustering. We formulate our problem as an instance of large margin structured prediction and prove that it can be solved very efficiently in closed-form. The complexity of our method is (in most cases) linear in the size of the training dataset. We further reveal a striking similarity between our approach and multivariate linear regression. Experiments on both synthetic and real datasets confirm several orders of magnitude speedup while still achieving state-of-the-art performance.

While deep convolutional neural networks (CNNs) have shown a great success in single-label image classification, it is important to note that most real world images contain multiple labels, which could correspond to different objects, scenes, actions and attributes in an image. Traditional approaches to multi-label image classification learn independent classifiers for each category and employ ranking or thresholding on the classification results. These techniques, although working well, fail to explicitly exploit the label dependencies in an image. In this paper, we utilize recurrent neural networks (RNNs) to address this problem. Combined with CNNs, the proposed CNN-RNN framework learns a joint image-label embedding to characterize the semantic label dependency as well as the image-label relevance, and it can be trained end-to-end from scratch to integrate both information in an unified framework. Experimental results on public benchmark datasets demonstrate that the proposed architecture achieves better performance than the state-of-the-art multi-label classification models.

Given an image of a handwritten word, a CNN is employed to estimate its n-gram frequency profile, which is the set of n-grams contained in the word. Frequencies for unigrams, bigrams and trigrams are estimated for the entire word and for parts of it. Canonical Correlation Analysis is then used to match the estimated profile to the true profiles of all words in a large dictionary. The CNN that is used employs several novelties such as the use of multiple fully connected branches. Applied to all commonly used handwriting recognition benchmarks, our method outperforms, by a very large margin, all existing methods.

We present a simple and effective blind image deblurring method based on the dark channel prior. Our work is inspired by the interesting observation that the dark channel of blurred images is less sparse. While most image patches in the clean image contain some dark pixels, these pixels are not dark when averaged with neighboring high-intensity pixels during the blur process.Our analysis shows that this change in the sparsity of the dark channel is an inherent property of the blur process, both theoretically and empirically. This change in the sparsity of the dark channel is an inherent property of the blur process, which we both prove mathematically and validate using training data. Therefore, enforcing the sparsity of the dark channel helps blind deblurring on various scenarios, including natural, face, text, and low-illumination images. However, sparsity of the dark channel introduces a non-convex non-linear optimization problem. We introduce a linear approximation of the min operator to compute the dark channel. Our look-up-table-based method converges fast in practice and can be directly extended to non-uniform deblurring. Extensive experiments show that our method achieves state-of-the-art results on deblurring natural images and compares favorably methods that are well-engineered for specific scenarios.

We present a highly accurate single image superresolution (SR) method. Our method uses a very deep convolutional network inspired by VGG-net used for ImageNet classification [19]. We find increasing our network depth shows a significant improvement in accuracy. Our final model uses 20 weight layers. By cascading small filters many times in a deep network structure, contextual information over large image regions is exploited in an efficient way. With very deep networks, however, convergence speed becomes a critical issue during training. We propose a simple yet effective training procedure. We learn residuals only and use extremely high learning rates (104 times higher than SRCNN [6]) enabled by adjustable gradient clipping. Our proposed method performs better than existing methods in accuracy and visual improvements in our results are easily noticeable.