ICCV 2015 论文列表

We describe a technique to recover depth from a light field (LF) using two proposed features of the LF focal stack. One feature is the property that non-occluding pixels exhibit symmetry along the focal depth dimension centered at the in-focus slice. The other is a data consistency measure based on analysis-by-synthesis, i.e., the difference between the synthesized focal stack given the hypothesized depth map and that from the LF. These terms are used in an iterative optimization framework to extract scene depth. Experimental results on real Lytro and Raytrix data demonstrate that our technique outperforms state-of-the-art solutions and is significantly more robust to noise and under-sampling.

In this paper, we present a joint iterative anaglyph stereo matching and colorization framework for obtaining a set of disparity maps and colorized images. Conventional stereo matching algorithms fail when addressing anaglyph images that do not have similar intensities on their two respective view images. To resolve this problem, we propose two novel data costs using local color prior and reverse intensity distribution factor for obtaining accurate depth maps. To colorize an anaglyph image, each pixel in one view is warped to another view using the obtained disparity values of non-occluded regions. A colorization algorithm using optimization is then employed with additional constraint to colorize the remaining occluded regions. Experimental results confirm that the proposed unified framework is robust and produces accurate depth maps and colorized stereo images.

Determining dense semantic correspondences across objects and scenes is a difficult problem that underpins many higher-level computer vision algorithms. Unlike canonical dense correspondence problems which consider images that are spatially or temporally adjacent, semantic correspondence is characterized by images that share similar high-level structures whose exact appearance and geometry may differ. Motivated by object recognition literature and recent work on rapidly estimating linear classifiers, we treat semantic correspondence as a constrained detection problem, where an exemplar LDA classifier is learned for each pixel. LDA classifiers have two distinct benefits: (i) they exhibit higher average precision than similarity metrics typically used in correspondence problems, and (ii) unlike exemplar SVM, can output globally interpretable posterior probabilities without calibration, whilst also being significantly faster to train. We pose the correspondence problem as a graphical model, where the unary potentials are computed via convolution with the set of exemplar classifiers, and the joint potentials enforce smoothly varying correspondence assignment.

Given a scene, what is going to move, and in what direction will it move? Such a question could be considered a non-semantic form of action prediction. In this work, we present a convolutional neural network (CNN) based approach for motion prediction. Given a static image, this CNN predicts the future motion of each and every pixel in the image in terms of optical flow. Our CNN model leverages the data in tens of thousands of realistic videos to train our model. Our method relies on absolutely no human labeling and is able to predict motion based on the context of the scene. Because our CNN model makes no assumptions about the underlying scene, it can predict future optical flow on a diverse set of scenarios. We outperform all previous approaches by large margins.

Deformable surface tracking from monocular images is well-known to be under-constrained. Occlusions often make the task even more challenging, and can result in failure if the surface is not sufficiently textured. In this work, we explicitly address the problem of 3D reconstruction of poorly textured, occluded surfaces, proposing a framework based on a template-matching approach that scales dense robust features by a relevancy score. Our approach is extensively compared to current methods employing both local feature matching and dense template alignment. We test on standard datasets as well as on a new dataset (that will be made publicly available) of a sparsely textured, occluded surface. Our framework achieves state-of-the-art results for both well and poorly textured, occluded surfaces.

We propose a dense continuous-time tracking and mapping method for RGB-D cameras. We parametrize the camera trajectory using continuous B-splines and optimize the trajectory through dense, direct image alignment. Our method also directly models rolling shutter in both RGB and depth images within the optimization, which improves tracking and reconstruction quality for low-cost CMOS sensors. Using a continuous trajectory representation has a number of advantages over a discrete-time representation (e.g. camera poses at the frame interval). With splines, less variables need to be optimized than with a discrete representation, since the trajectory can be represented with fewer control points than frames. Splines also naturally include smoothness constraints on derivatives of the trajectory estimate. Finally, the continuous trajectory representation allows to compensate for rolling shutter effects, since a pose estimate is available at any exposure time of an image. Our approach demonstrates superior quality in tracking and reconstruction compared to approaches with discrete-time or global shutter assumptions.

Rectified activation units (rectifiers) are essential for state-of-the-art neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on the learnable activation and advanced initialization, we achieve 4.94% top-5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66%). To our knowledge, our result is the first to surpass the reported human-level performance (5.1%) on this dataset.

Capturing the 3D motion of dynamic, non-rigid objects has attracted significant attention in computer vision. Existing methods typically require either complete 3D volumetric observations, or a shape template. In this paper, we introduce a template-less 4D reconstruction method that incrementally fuses highly-incomplete 3D observations of a deforming object, and generates a complete, temporally-coherent shape representation of the object. To this end, we design an online algorithm that alternatively registers new observations to the current model estimate and updates the model. We demonstrate the effectiveness of our approach at reconstructing non-rigidly moving objects from highly-incomplete measurements on both sequences of partial 3D point clouds and Kinect videos.

In this paper we evaluate the quality of the activation layers of a convolutional neural network (CNN) for the generation of object proposals. We generate hypotheses in a sliding-window fashion over different activation layers and show that the final convolutional layers can find the object of interest with high recall but poor localization due to the coarseness of the feature maps. Instead, the first layers of the network can better localize the object of interest but with a reduced recall. Based on this observation we design a method for proposing object locations that is based on CNN features and that combines the best of both worlds. We build an inverse cascade that, going from the final to the initial convolutional layers of the CNN, selects the most promising object locations and refines their boxes in a coarse-to-fine manner. The method is efficient, because i) it uses the same features extracted for detection, ii) it aggregates features using integral images, and iii) it avoids a dense evaluation of the proposals due to the inverse coarse-to-fine cascade. The method is also accurate; it outperforms most of the previously proposed object proposals approaches and when plugged into a CNN-based detector produces state-of-the-art detection performance.

Today, there are two major paradigms for vision-based autonomous driving systems: mediated perception approaches that parse an entire scene to make a driving decision, and behavior reflex approaches that directly map an input image to a driving action by a regressor. In this paper, we propose a third paradigm: a direct perception approach to estimate the affordance for driving. We propose to map an input image to a small number of key perception indicators that directly relate to the affordance of a road/traffic state for driving. Our representation provides a set of compact yet complete descriptions of the scene to enable a simple controller to drive autonomously. Falling in between the two extremes of mediated perception and behavior reflex, we argue that our direct perception representation provides the right level of abstraction. To demonstrate this, we train a deep Convolutional Neural Network using recording from 12 hours of human driving in a video game and show that our model can work well to drive a car in a very diverse set of virtual environments. We also train a model for car distance estimation on the KITTI dataset. Results show that our direct perception approach can generalize well to real driving images. Source code and data are available on our project website.

Existing object proposal approaches use primarily bottom-up cues to rank proposals, while we believe that "objectness" is in fact a high level construct. We argue for a data-driven, semantic approach for ranking object proposals. Our framework, which we call DeepBox, uses convolutional neural networks (CNNs) to rerank proposals from a bottom-up method. We use a novel four-layer CNN architecture that is as good as much larger networks on the task of evaluating objectness while being much faster. We show that DeepBox significantly improves over the bottom-up ranking, achieving the same recall with 500 proposals as achieved by bottom-up methods with 2000. This improvement generalizes to categories the CNN has never seen before and leads to a 4.5-point gain in detection mAP. Our implementation achieves this performance while running at 260 ms per image.

We present Deep Neural Decision Forests - a novel approach that unifies classification trees with the representation learning functionality known from deep convolutional networks, by training them in an end-to-end manner. To combine these two worlds, we introduce a stochastic and differentiable decision tree model, which steers the representation learning usually conducted in the initial layers of a (deep) convolutional network. Our model differs from conventional deep networks because a decision forest provides the final predictions and it differs from conventional decision forests since we propose a principled, joint and global optimization of split and leaf node parameters. We show experimental results on benchmark machine learning datasets like MNIST and ImageNet and find on-par or superior results when compared to state-of-the-art deep models. Most remarkably, we obtain Top5-Errors of only 7.84%/6.38% on ImageNet validation data when integrating our forests in a single-crop, single/seven model GoogLeNet architecture, respectively. Thus, even without any form of training data set augmentation we are improving on the 6.67% error obtained by the best GoogLeNet architecture (7 models, 144 crops).

Deep learning techniques have been successfully applied in many areas of computer vision, including low-level image restoration problems. For image super-resolution, several models based on deep neural networks have been recently proposed and attained superior performance that overshadows all previous handcrafted models. The question then arises whether large-capacity and data-driven models have become the dominant solution to the ill-posed super-resolution problem. In this paper, we argue that domain expertise represented by the conventional sparse coding model is still valuable, and it can be combined with the key ingredients of deep learning to achieve further improved results. We show that a sparse coding model particularly designed for super-resolution can be incarnated as a neural network, and trained in a cascaded structure from end to end. The interpretation of the network based on sparse coding leads to much more efficient and effective training, as well as a reduced model size. Our model is evaluated on a wide range of images, and shows clear advantage over existing state-of-the-art methods in terms of both restoration accuracy and human subjective quality.

This paper investigates problems of image style, aesthetics, and quality estimation, which require fine-grained details from high-resolution images, utilizing deep neural network training approach. Existing deep convolutional neural networks mostly extracted one patch such as a down-sized crop from each image as a training example. However, one patch may not always well represent the entire image, which may cause ambiguity during training. We propose a deep multi-patch aggregation network training approach, which allows us to train models using multiple patches generated from one image. We achieve this by constructing multiple, shared columns in the neural network and feeding multiple patches to each of the columns. More importantly, we propose two novel network layers (statistics and sorting) to support aggregation of those patches. The proposed deep multi-patch aggregation network integrates shared feature learning and aggregation function learning into a unified framework. We demonstrate the effectiveness of the deep multi-patch aggregation network on the three problems, i.e., image style recognition, aesthetic quality categorization, and image quality estimation. Our models trained using the proposed networks significantly outperformed the state of the art in all three applications.

Recent advances in pedestrian detection are attained by transferring the learned features of Convolutional Neural Network (ConvNet) to pedestrians. This ConvNet is typically pre-trained with massive general object categories (e.g. ImageNet). Although these features are able to handle variations such as poses, viewpoints, and lightings, they may fail when pedestrian images with complex occlusions are present. Occlusion handling is one of the most important problem in pedestrian detection. Unlike previous deep models that directly learned a single detector for pedestrian detection, we propose DeepParts, which consists of extensive part detectors. DeepParts has several appealing properties. First, DeepParts can be trained on weakly labeled data, i.e. only pedestrian bounding boxes without part annotations are provided. Second, DeepParts is able to handle low IoU positive proposals that shift away from ground truth. Third, each part detector in DeepParts is a strong detector that can detect pedestrian by observing only a part of a proposal. Extensive experiments in Caltech dataset demonstrate the effectiveness of DeepParts, which yields a new state-of-the-art miss rate of 11:89%, outperforming the second best method by 10%.

Predicting face attributes in the wild is challenging due to complex face variations. We propose a novel deep learning framework for attribute prediction in the wild. It cascades two CNNs, LNet and ANet, which are fine-tuned jointly with attribute tags, but pre-trained differently. LNet is pre-trained by massive general object categories for face localization, while ANet is pre-trained by massive face identities for attribute prediction. This framework not only outperforms the state-of-the-art with a large margin, but also reveals valuable facts on learning face representation. (1) It shows how the performances of face localization (LNet) and attribute prediction (ANet) can be improved by different pre-training strategies. (2) It reveals that although the filters of LNet are fine-tuned only with image-level attribute tags, their response maps over entire images have strong indication of face locations. This fact enables training LNet for face localization with only image-level annotations, but without face bounding boxes or landmarks, which are required by all attribute recognition works. (3) It also demonstrates that the high-level hidden neurons of ANet automatically discover semantic concepts after pre-training with massive face identities, and such concepts are significantly enriched after fine-tuning with attribute tags. Each attribute can be well explained with a sparse linear combination of these concepts.

The fully connected layers of a deep convolutional neural network typically contain over 90% of the network parameters, and consume the majority of the memory required to store the network. Reducing the number of parameters while preserving predictive performance is critically important for deploying deep neural networks in memory constrained environments such as GPUs or embedded devices. In this paper we show how kernel methods, in particular a single Fastfood layer, can be used to replace the fully connected layers in a deep convolutional neural network. This deep fried network is end-to-end trainable in conjunction with convolutional layers. Our new architecture substantially reduces the memory footprint of convolutional networks trained on MNIST and ImageNet with no drop in predictive performance

This paper investigates into the colorization problem which converts a grayscale image to a colorful version. This is a very difficult problem and normally requires manual adjustment to achieve artifact-free quality. For instance, it normally requires human-labelled color scribbles on the grayscale target image or a careful selection of colorful reference images (e.g., capturing the same scene in the grayscale target image). Unlike the previous methods, this paper aims at a high-quality fully-automatic colorization method. With the assumption of a perfect patch matching technique, the use of an extremely large-scale reference database (that contains sufficient color images) is the most reliable solution to the colorization problem. However, patch matching noise will increase with respect to the size of the reference database in practice. Inspired by the recent success in deep learning techniques which provide amazing modeling of large-scale data, this paper re-formulates the colorization problem so that deep learning techniques can be directly employed. To ensure artifact-free quality, a joint bilateral filtering based post-processing step is proposed. Numerous experiments demonstrate that our method outperforms the state-of-art algorithms both in terms of quality and speed.

Exploiting multiple modalities for semantic scene parsing has been shown to improve accuracy over the single modality scenario. Existing methods, however, assume that corresponding regions in two modalities have the same label. In this paper, we address the problem of data misalignment and label inconsistencies, e.g., due to moving objects, in semantic labeling, which violate the assumption of existing techniques. To this end, we formulate multimodal semantic labeling as inference in a CRF, and introduce latent nodes to explicitly model inconsistencies between two domains. These latent nodes allow us not only to leverage information from both domains to improve their labeling, but also to cut the edges between inconsistent regions. To eliminate the need for hand tuning the parameters of our model, we propose to learn intra-domain and inter-domain potential functions from training data. We demonstrate the benefits of our approach on two publicly available datasets containing 2D imagery and 3D point clouds. Thanks to our latent nodes and our learning strategy, our method outperforms the state-of-the-art in both cases.