CVPR 2015 Papers

Arguably the most common cause of image degradation is compression. This papers presents a novel approach to restoring JPEG-compressed images. The main innovation is in the approach of exploiting residual redundancies of JPEG code streams and sparsity properties of latent images. The restoration is a sparse coding process carried out jointly in the DCT and pixel domains. The prowess of the proposed approach is directly restoring DCT coefficients of the latent image to prevent the spreading of quantization errors into the pixel domain, and at the same time using on-line machine-learnt local spatial features to regulate the solution of the underlying inverse problem. Experimental results are encouraging and show the promise of the new approach in significantly improving the quality of DCT-coded images.

Depth maps captured by consumer-level depth cameras such as Kinect are usually degraded by noise, missing values, and quantization. In this paper, we present a data-driven approach for refining degraded RAW depth maps that are coupled with an RGB image. The key idea of our approach is to take advantage of a training set of high-quality depth data and transfer its information to the RAW depth map through multi-scale dictionary learning. Utilizing a sparse representation, our method learns a dictionary of geometric primitives which captures the correlation between high-quality mesh data, RAW depth maps and RGB images. The dictionary is learned and applied in a manner that accounts for various practical issues that arise in dictionary-based depth refinement. Compared to previous approaches that only utilize the correlation between RAW depth maps and RGB images, our method produces improved depth maps without over-smoothing. Since our approach is data driven, the refinement can be targeted to a specific class of objects by employing a corresponding training set. In our experiments, we show that this leads to additional improvements in recovering depth maps of human faces.

Despite the great progress achieved in recognizing objects as 2D bounding boxes in images, it is still very challenging to detect occluded objects and estimate the 3D properties of multiple objects from a single image. In this paper, we propose a novel object representation, 3D Voxel Pattern (3DVP), that jointly encodes the key properties of objects including appearance, 3D shape, viewpoint, occlusion and truncation. We discover 3DVPs in a data-driven way, and train a bank of specialized detectors for a dictionary of 3DVPs. The 3DVP detectors are capable of detecting objects with specific visibility patterns and transferring the meta-data from the 3DVPs to the detected objects, such as 2D segmentation mask, 3D pose as well as occlusion or truncation boundaries. The transferred meta-data allows us to infer the occlusion relationship among objects, which in turn provides improved object recognition results. Experiments are conducted on the KITTI detection benchmark and the outdoor-scene dataset. We improve state-of-the-art results on car detection and pose estimation with notable margins (6% in difficult data of KITTI). We also verify the ability of our method in accurately segmenting objects from the background and localizing them in 3D.

Most of the approaches for discovering visual attributes in images demand significant supervision, which is cumbersome to obtain. In this paper, we aim to discover visual attributes in a weakly supervised setting that is commonly encountered with contemporary image search engines. For instance, given a noun (say forest) and its associated attributes (say dense, sunlit, autumn), search engines can now generate many valid images for any attribute-noun pair (dense forests, autumn forests, etc). However, images for an attribute-noun pair do not contain any information about other attributes (like which forests in the autumn are dense too). Thus, a weakly supervised scenario occurs: each of the M attributes corresponds to a class such that a training image in class m = 1,...,M contains a single label that indicates the presence of the m(th) attribute only. The task is to discover all the attributes present in a test image. Deep Convolutional Neural Networks (CNNs) have enjoyed remarkable success in vision applications recently. However, in a weakly supervised scenario, widely used CNN training procedures do not learn a robust model for predicting multiple attribute labels simultaneously. The primary reason is that the attributes highly co-occur within the training data, and unlike objects, do not generally exist as well-defined spatial boundaries within the image. To ameliorate this limitation, we propose Deep-Carving, a novel training procedure with CNNs, that helps the net efficiently carve itself for the task of multiple attribute prediction. During training, the responses of the feature maps are exploited in an ingenious way to provide the net with multiple pseudo-labels (for training images) for subsequent iterations. The process is repeated periodically after a fixed number of iterations, and enables the net carve itself iteratively for efficiently disentangling features. Additionally, we contribute a noun-adjective pairing inspired Natural Scenes Attributes Dataset to the research community, CAMIT - NSAD, containing a number of co-occurring attributes within a noun category. We describe, in detail, salient aspects of this dataset. Our experiments on CAMIT-NSAD and the SUN Attributes Dataset, with weak supervision, clearly demonstrate that the Deep-Carved CNNs consistently achieve considerable improvement in the precision of attribute prediction over popular baseline methods.

Establishing dense visual correspondence between multiple images is a fundamental task in many applications of computer vision and computational photography. Classical approaches, which aim to estimate dense stereo and optical flow fields for images adjacent in viewpoint or in time, have been dramatically advanced in recent studies. However, finding reliable visual correspondence in multi-modal or multi-spectral images still remains unsolved. In this paper, we propose a new dense matching descriptor, called dense adaptive self-correlation (DASC), to effectively address this kind of matching scenarios. Based on the observation that a self-similarity existing within images is less sensitive to modality variations, we define a novel descriptor with a series of an adaptive self-correlation similarity for patches within a local support window. To further improve the matching quality and runtime efficiency, we propose a patch-wise receptive field pooling, in which a sampling pattern is optimized with a discriminative learning. Moreover, the computational redundancy that arises when computing densely sampled descriptor over an entire image is dramatically reduced by applying fast edge-aware filtering. Experiments demonstrate the outstanding performance of the DASC descriptor in many cases of multi-modal and multi-spectral correspondence.

Sharing information between multiple tasks enables algorithms to achieve good generalization performance even from small amounts of training data. However, in a realistic scenario of multi-task learning not all tasks are equally related to each other, hence it could be advantageous to transfer information only between the most related tasks. In this work we propose an approach that processes multiple tasks in a sequence with sharing between subsequent tasks instead of solving all tasks jointly. Subsequently, we address the question of curriculum learning of tasks, i.e. finding the best order of tasks to be learned. Our approach is based on a generalization bound criterion for choosing the task order that optimizes the average expected classification performance over all tasks. Our experimental results show that learning multiple related tasks sequentially can be more effective than learning them jointly, the order in which tasks are being solved affects the overall performance, and that our model is able to automatically discover a favourable order of tasks.

Cross-scene crowd counting is a challenging task where no laborious data annotation is required for counting people in new target surveillance crowd scenes unseen in the training set. The performance of most existing crowd counting methods drops significantly when they are applied to an unseen scene. To address this problem, we propose a deep convolutional neural network (CNN) for crowd counting, and it is trained alternatively with two related learning objectives, crowd density and crowd count. This proposed switchable learning approach is able to obtain better local optimum for both objectives. To handle an unseen target crowd scene, we present a data-driven method to fine-tune the trained CNN model for the target scene. A new dataset including 108 crowd scenes with nearly 200,000 head annotations is introduced to better evaluate the accuracy of cross-scene crowd counting methods. Extensive experiments on the proposed and another two existing datasets demonstrate the effectiveness and reliability of our approach.

In this paper we present a novel framework for cross-age face verification (FV) by seeking help from its ``competitor" named cross-face age verification (AV), i.e., deciding whether two face photos are taken at similar ages. While FV and AV share some common features, FV pursues age insensitivity and AV seeks age sensitivity. Such correlation suggests that AV may be used to guide feature selection in FV, i.e., by reducing the chance of choosing age sensitive features. Driven by this intuition, we propose to learn a solution for cross-age face verification by coordinating with a solution for age verification. Specifically, a joint additive model is devised to simultaneously handling both tasks, while encoding feature coordination by a competition regularization term. Then, an alternating greedy coordinate descent (AGCD) algorithm is developed to solve this joint model. As shown in our experiments, the algorithm effectively balances feature sharing and feature exclusion between the two tasks; and, for face verification, the algorithm effectively removes distracting features used in age verification. To evaluate the proposed algorithm, we conduct cross-age face verification experiments using two benchmark cross-age face datasets, FG-Net and MORPH. In all experiments, our algorithm achieves very promising results and outperforms all previously tested solutions.

Correlation filters take advantage of specific properties in the Fourier domain allowing them to be estimated efficiently: O(ND log D) in the frequency domain, versus O(D^3 + ND^2) spatially where D is signal length, and N is the number of signals. Recent extensions to correlation filters, such as MOSSE, have reignited interest of their use in the vision community due to their robustness and attractive computational properties. In this paper we demonstrate, however, that this computational efficiency comes at a cost. Specifically, we demonstrate that only 1/D proportion of shifted examples are unaffected by boundary effects which has a dramatic effect on detection/tracking performance. In this paper, we propose a novel approach to correlation filter estimation that: (i) takes advantage of inherent computational redundancies in the frequency domain, (ii) dramatically reduces boundary effects, and (iii) is able to implicitly exploit all possible patches densely extracted from training examples during learning process. Impressive object tracking and detection results are presented in terms of both accuracy and computational efficiency.

Though recent advanced convolutional neural networks (CNNs) have been improving the image recognition accuracy, the models are getting more complex and time-consuming. For real-world applications in industrial and commercial scenarios, engineers and developers are often faced with the requirement of constrained time budget. In this paper, we investigate the accuracy of CNNs under constrained time cost. Under this constraint, the designs of the network architectures should exhibit as trade-offs among the factors like depth, numbers of filters, filter sizes, etc. With a series of controlled comparisons, we progressively modify a baseline model while preserving its time complexity. This is also helpful for understanding the importance of the factors in network designs. We present an architecture that achieves very competitive accuracy in the ImageNet dataset (11.8% top-5 error, 10-view test), yet is 20% faster than ``AlexNet'' (16.0% top-5 error, 10-view test).

The topic of semantic segmentation has witnessed considerable progress due to the powerful features learned by convolutional neural networks (CNNs). The current leading approaches for semantic segmentation exploit shape information by extracting CNN features from masked image regions. This strategy introduces artificial boundaries on the images and may impact the quality of the extracted features. Besides, the operations on the raw image domain require to compute thousands of networks on a single image, which is time-consuming. In this paper, we propose to exploit shape information via masking convolutional features. The proposal segments (e.g., super-pixels) are treated as masks on the convolutional feature maps. The CNN features of segments are directly masked out from these maps and used to train classifiers for recognition. We further propose a joint method to handle objects and "stuff" (e.g., grass, sky, water) in the same framework. State-of-the-art results are demonstrated on the challenging PASCAL VOC benchmarks, with a compelling computational speed.

In this work, we propose a new type of visibility measurement named Continuous Visibility Feature (CVF). We say that a point q on the mesh is continuously visible from another point p if there exists a geodesic path connecting p and q that is entirely visible by p. In order to efficiently estimate the continuous visibility for all the vertices in a model, we propose two approaches that use specific CVF properties to avoid exhaustive visibility tests. CVF is then measured as the area of the continuously visible region. With this stronger visibility measure, we show that CVF better encodes the surface and part information of mesh than the tradition line-of-sight based visibility. For example, we show that existing segmentation algorithms can generate better segmentation results using CVF and its variants than using other visibility-based shape descriptors, such as shape diameter function. Similar to visibility and other mesh surface features, continuous visibility would have many applications.

While humans can easily separate unknown objects into meaningful parts, recent segmentation methods can only achieve similar partitionings by training on human-annotated ground-truth data. Here we introduce a bottom-up method for segmenting 3D point clouds into functional parts which does not require supervision and achieves equally good results. Our method uses local concavities as an indicator for inter-part boundaries. We show that this criterion is efficient to compute and generalizes well across different object classes. The algorithm employs a novel locally constrained geometrical boundary model which proposes greedy cuts through a local concavity graph. Only planar cuts are considered and evaluated using a cost function, which rewards cuts orthogonal to concave edges. Additionally, a local clustering constraint is applied to ensure the partitioning only affects relevant locally concave regions. We evaluate our algorithm on recordings from an RGB-D camera as well as the Princeton Segmentation Benchmark, using a fixed set of parameters across all object classes. This stands in stark contrast to most reported results which require either knowing the number of parts or annotated ground-truth for learning. Our approach outperforms all existing bottom-up methods (reducing the gap to human performance by up to 50%) and achieves scores similar to top-down data-driven approaches.

Discovering visual knowledge from weakly labeled data is crucial to scale up computer vision recognition systems, since it is expensive to obtain fully labeled data for a large number of concept categories. In this paper, we propose ConceptLearner, which is a scalable approach to discover visual concepts from weakly labeled image collections. Thousands of visual concept detectors are learned automatically, without human in the loop for additional annotation. We show that these learned detectors could be applied to recognize concepts at image-level and to detect concepts at image region-level accurately. Under domain-specific supervision, we further evaluate the learned concepts for scene recognition on SUN database and for object detection on Pascal VOC 2007. ConceptLearner shows promising performance compared to fully supervised and weakly supervised methods.

We present a method for extracting depth information from a rectified image pair. We train a convolutional neural network to predict how well two image patches match and use it to compute the stereo matching cost. The cost is refined by cross-based cost aggregation and semiglobal matching, followed by a left-right consistency check to eliminate errors in the occluded regions. Our stereo method achieves an error rate of 2.61 % on the KITTI stereo dataset and is currently (August 2014) the top performing method on this dataset.

We propose a novel solution for computing the relative pose between two generalized cameras that includes reconciling the internal scale of the generalized cameras. This approach can be used to compute a similarity transformation between two coordinate systems, making it useful for loop closure in visual odometry and registering multiple structure from motion reconstructions together. In contrast to alternative similarity transformation methods, our approach uses 2D-2D image correspondences thus is not subject to the depth uncertainty that often arises with 3D points. We utilize a known vertical direction (which may be easily obtained from IMU data or vertical vanishing point detection) of the generalized cameras to solve the generalized relative pose and scale problem as an efficient Quadratic Eigenvalue Problem. To our knowledge, this is the first method for computing similarity transformations that does not require any 3D information. Our experiments on synthetic and real data demonstrate that this leads to improved performance compared to methods that use 3D-3D or 2D-3D correspondences, especially as the depth of the scene increases.

The increase in size of large image databases makes the problem of efficient retrieval extremely challenging. This is especially true in the case of high dimensional data where even operations like hashing become expensive because of costly projection operators. Unlike most hashing methods that sacrifice accuracy for speed, we propose a novel method that improves the speed of high dimensional image retrieval by several orders of magnitude without any significant drop in performance. To do this, we propose to learn computationally bounded sparse projections for the encoding step. To further increase the accuracy of the method, we add an orthogonality constraint on projections to reduce bit correlation. We then introduce an iterative scheme that jointly optimizes this objective, which helps us obtain fast and efficient projections. We demonstrate this technique on large retrieval databases, specifically ImageNET, GIST1M and SUN-attribute for the task of nearest neighbor retrieval, and show that our method achieves a speed-up of up to a factor of 100 over state-of-the-art methods, while having on-par and in some cases even better accuracy.

Distance metric plays a key role in grouping superpixels to produce object proposals for object detection. We observe that existing distance metrics work primarily for low complexity cases. In this paper, we develop a novel distance metric for grouping two superpixels in high-complexity scenarios. Combining them, a complexity-adaptive distance measure is produced that achieves improved grouping in different levels of complexity. Our extensive experimentation shows that our method can achieve good results in the PASCAL VOC 2012 dataset surpassing the latest state-of-the-art methods.

Our goal is to recover a complete 3D model from a depth image of an object. Existing approaches rely on user interaction or apply to a limited class of objects, such as chairs. We aim to fully automatically reconstruct a 3D model from any category. We take an exemplar-based approach: retrieve similar objects in a database of 3D models using view-based matching and transfer the symmetries and surfaces from retrieved models. We investigate completion of 3D models in three cases: novel view (model in database); novel model (models for other objects of the same category in database); and novel category (no models from the category in database).