CVPR 2018 Papers

Semantic concept hierarchy is yet under-explored for semantic segmentation due to the inefficiency and complicated optimization of incorporating structural inference into the dense prediction. This lack of modeling dependencies among concepts severely limits the generalization capability of segmentation models for open set large-scale vocabularies. Prior works thus must tune highly-specified models for each task due to the label discrepancy across datasets. In this paper, we propose a Dynamic-Structured Semantic Propagation Network (DSSPN) that builds a semantic neuron graph to explicitly incorporate the concept hierarchy into dynamic network construction, leading to an interpretable reasoning process. Each neuron for one super-class (eg food) represents the instantiated module for recognizing among fine-grained child concepts (eg editable fruit or pizza), and then its learned features flow into the child neurons (eg distinguishing between orange or apple) for hierarchical categorization in finer levels. A dense semantic-enhanced neural block propagates the learned knowledge of all ancestral neurons into each fine-grained child neuron for progressive feature evolving. During training, DSSPN performs the dynamic-structured neuron computational graph by only activating a sub-graph of neurons for each image. Another merit of such semantic explainable structure is the ability to learn a unified model concurrently on diverse datasets by selectively activating different neuron sub-graphs for each annotation at each step. Extensive experiments on four public semantic segmentation datasets (i.e. ADE20K, COCO-Stuff, Cityscape and Mapillary) demonstrate the superiority of DSSPN, and a universal segmentation model that is jointly trained on diverse datasets can surpass the common fine-tuning scheme for exploiting multi-domain knowledge.

We introduce a generic framework that reduces the computational cost of object detection while retaining accuracy for scenarios where objects with varied sizes appear in high resolution images. Detection progresses in a coarse-to-fine manner, first on a down-sampled version of the image and then on a sequence of higher resolution regions identified as likely to improve the detection accuracy. Built upon reinforcement learning, our approach consists of a model (R-net) that uses coarse detection results to predict the potential accuracy gain for analyzing a region at a higher resolution and another model (Q-net) that sequentially selects regions to zoom in. Experiments on the Caltech Pedestrians dataset show that our approach reduces the number of processed pixels by over 50% without a drop in detection accuracy. The merits of our approach become more significant on a high resolution test set collected from YFCC100M dataset, where our approach maintains high detection performance while reducing the number of processed pixels by about 70% and the detection time by over 50%.

In this paper, we present a detailed design of dynamic video segmentation network (DVSNet) for fast and efficient semantic video segmentation. DVSNet consists of two convolutional neural networks: a segmentation network and a flow network. The former generates highly accurate semantic segmentations, but is deeper and slower. The latter is much faster than the former, but its output requires further processing to generate less accurate semantic segmentations. We explore the use of a decision network to adaptively assign different frame regions to different networks based on a metric called expected confidence score. Frame regions with a higher expected confidence score traverse the flow network. Frame regions with a lower expected confidence score have to pass through the segmentation network. We have extensively performed experiments on various configurations of DVSNet, and investigated a number of variants for the proposed decision network. The experimental results show that our DVSNet is able to achieve up to 70.4% mIoU at 19.8 fps on the Cityscape dataset. A high speed version of DVSNet is able to deliver an fps of 30.4 with 63.2% mIoU on the same dataset. DVSNet is also able to reduce up to 95% of the computational workloads.

Due to the spatially variant blur caused by camera shake and object motions under different scene depths, deblurring images captured from dynamic scenes is challenging. Although recent works based on deep neural networks have shown great progress on this problem, their models are usually large and computationally expensive. In this paper, we propose a novel spatially variant neural network to address the problem. The proposed network is composed of three deep convolutional neural networks (CNNs) and a recurrent neural network (RNN). RNN is used as a deconvolution operator performed on feature maps extracted from the input image by one of the CNNs. Another CNN is used to learn the weights for the RNN at every location. As a result, the RNN is spatially variant and could implicitly model the deblurring process with spatially variant kernels. The third CNN is used to reconstruct the final deblurred feature maps into restored image. The whole network is end-to-end trainable. Our analysis shows that the proposed network has a large receptive field even with a small model size. Quantitative and qualitative evaluations on public datasets demonstrate that the proposed method performs favorably against state-of-the-art algorithms in terms of accuracy, speed, and model size.

In this work, we introduce a new algorithm for analyzing a diagram, which contains visual and textual information in an abstract and integrated way. Whereas diagrams contain richer information compared with individual image-based or language-based data, proper solutions for automatically understanding them have not been proposed due to their innate characteristics of multi-modality and arbitrariness of layouts. To tackle this problem, we propose a unified diagram-parsing network for generating knowledge from diagrams based on an object detector and a recurrent neural network designed for a graphical structure. Specifically, we propose a dynamic graph-generation network that is based on dynamic memory and graph theory. We explore the dynamics of information in a diagram with activation of gates in gated recurrent unit (GRU) cells. On publicly available diagram datasets, our model demonstrates a state-of-the-art result that outperforms other baselines. Moreover, further experiments on question answering shows potentials of the proposed method for various applications.

The human visual system has the remarkably ability to be able to effortlessly learn novel concepts from only a few examples. Mimicking the same behavior on machine learning vision systems is an interesting and very challenging research problem with many practical advantages on real world vision applications. In this context, the goal of our work is to devise a few-shot visual learning system that during test time it will be able to efficiently learn novel categories from only a few training data while at the same time it will not forget the initial categories on which it was trained (here called base categories). To achieve that goal we propose (a) to extend an object recognition system with an attention based few-shot classification weight generator, and (b) to redesign the classifier of a ConvNet model as the cosine similarity function between feature representations and classification weight vectors. The latter, apart from unifying the recognition of both novel and base categories, it also leads to feature representations that generalize better on "unseen" categories. We extensively evaluate our approach on Mini-ImageNet where we manage to improve the prior state-of-the-art on few-shot recognition (i.e., we achieve 56.20% and 73.00% on the 1-shot and 5-shot settings respectively) while at the same time we do not sacrifice any accuracy on the base categories, which is a characteristic that most prior approaches lack. Finally, we apply our approach on the recently introduced few-shot benchmark of Bharath and Girshick where we also achieve state-of-the-art results.

Partial face recognition (PFR) in unconstrained environment is a very important task, especially in video surveillance, mobile devices, etc. However, a few studies have tackled how to recognize an arbitrary patch of a face image. This study combines Fully Convolutional Network (FCN) with Sparse Representation Classification (SRC) to propose a novel partial face recognition approach, called Dynamic Feature Matching (DFM), to address partial face images regardless of sizes. Based on DFM, we propose a sliding loss to optimize FCN by reducing the intra-variation between a face patch and face images of a subject, which further improves the performance of DFM. The proposed DFM is evaluated on several partial face databases, including LFW, YTF and CASIA-NIR-Distance databases. Experimental results demonstrate the effectiveness and advantages of DFM in comparison with state-of-the-art PFR methods.

Domain adaptation attempts to transfer the knowledge obtained from the source domain to the target domain, i.e., the domain where the testing data are. The main challenge lies in the distribution discrepancy between source and target domain. Most existing works endeavor to learn domain invariant representation usually by minimizing a distribution distance, e.g., MMD and the discriminator in the recently proposed generative adversarial network (GAN). Following the similar idea of GAN, this work proposes a novel GAN architecture with duplex adversarial discriminators (referred to as DupGAN), which can achieve domain-invariant representation and domain transformation. Specifically, our proposed network consists of three parts, an encoder, a generator and two discriminators. The encoder embeds samples from both domains into the latent representation, and the generator decodes the latent representation to both source and target domains respectively conditioned on a domain code, i.e., achieves domain transformation. The generator is pitted against duplex discriminators, one for source domain and the other for target, to ensure the reality of domain transformation, the latent representation domain invariant and the category information of it preserved as well. Our proposed work achieves the state-of-the-art performance on unsupervised domain adaptation of digit classification and object recognition.

Inspired by the recent neuroscience studies on the left-right asymmetry of the human brain in processing low and high spatial frequency information, this paper introduces a dual skipping network which carries out coarse-to-fine object categorization. Such a network has two branches to simultaneously deal with both coarse and fine-grained classification tasks. Specifically, we propose a layer-skipping mechanism that learns a gating network to predict which layers to skip in the testing stage. This layer-skipping mechanism endows the network with good flexibility and capability in practice. Evaluations are conducted on several widely used coarse-to-fine object categorization benchmarks, and promising results are achieved by our proposed network model.

Typical person re-identification (ReID) methods usually describe each pedestrian with a single feature vector and match them in a task-specific metric space. However, the methods based on a single feature vector are not sufficient enough to overcome visual ambiguity, which frequently occurs in real scenario. In this paper, we propose a novel end-to-end trainable framework, called Dual ATtention Matching network (DuATM), to learn context-aware feature sequences and perform attentive sequence comparison simultaneously. The core component of our DuATM framework is a dual attention mechanism, in which both intra-sequence and inter-sequence attention strategies are used for feature refinement and feature-pair alignment, respectively. Thus, detailed visual cues contained in the intermediate feature sequences can be automatically exploited and properly compared. We train the proposed DuATM network as a siamese network via a triplet loss assisted with a de-correlation loss and a cross-entropy loss. We conduct extensive experiments on both image and video based ReID benchmark datasets. Experimental results demonstrate the significant advantages of our approach compared to the state-of-the-art methods.

We propose DoubleFusion, a new real-time system that combines volumetric dynamic reconstruction with data-driven template fitting to simultaneously reconstruct detailed geometry, non-rigid motion and the inner human body shape from a single depth camera. One of the key contributions of this method is a double layer representation consisting of a complete parametric body shape inside and a gradually fused outer surface layer. A pre-defined node graph on the body surface parameterizes the non-rigid deformations near the body and a free-form dynamically changing graph parameterizes the outer surface layer far from the body allowing more general reconstruction. We further propose a joint motion tracking method based on the double layer representation to enable robust and fast motion tracking performance. Moreover, the inner body shape is optimized online and forced to fit inside the outer surface layer. Overall, our method enables increasingly denoised, detailed and complete surface reconstructions, fast motion tracking performance and plausible inner body shape reconstruction in real-time. In particular, experiments show improved fast motion tracking and loop closure performance on more challenging scenarios.

A number of studies have found that today's Visual Question Answering (VQA) models are heavily driven by superficial correlations in the training data and lack sufficient image grounding. To encourage development of models geared towards the latter, we propose a new setting for VQA where for every question type, train and test sets have different prior distributions of answers. Specifically, we present new splits of the VQA v1 and VQA v2 datasets, which we call Visual Question Answering under Changing Priors (VQA-CP v1 and VQA-CP v2 respectively). First, we evaluate several existing VQA models under this new setting and show that their performance degrades significantly compared to the original VQA setting. Second, we propose a novel Grounded Visual Question Answering model (GVQA) that contains inductive biases and restrictions in the architecture specifically designed to prevent the model from 'cheating' by primarily relying on priors in the training data. Specifically, GVQA explicitly disentangles the recognition of visual concepts present in the image from the identification of plausible answer space for a given question, enabling the model to more robustly generalize across different distributions of answers. GVQA is built off an existing VQA model -- Stacked Attention Networks (SAN). Our experiments demonstrate that GVQA significantly outperforms SAN on both VQA-CP v1 and VQA-CP v2 datasets. Interestingly, it also outperforms more powerful VQA models such as Multimodal Compact Bilinear Pooling (MCB) in several cases. GVQA offers strengths complementary to SAN when trained and evaluated on the original VQA v1 and VQA v2 datasets. Finally, GVQA is more transparent and interpretable than existing VQA models.

In this paper, we tackle the problem of domain generalization: how to learn a generalized feature representation for an “unseen” target domain by taking the advantage of multiple seen source-domain data. We present a novel framework based on adversarial autoencoders to learn a generalized latent feature representation across domains for domain generalization. To be specific, we extend adversarial autoencoders by imposing the Maximum Mean Discrepancy (MMD) measure to align the distributions among different domains, and matching the aligned distribution to an arbitrary prior distribution via adversarial feature learning. In this way, the learned feature representation is supposed to be universal to the seen source domains because of the MMD regularization, and is expected to generalize well on the target domain because of the introduction of the prior distribution. We proposed an algorithm to jointly train different components of our proposed framework. Extensive experiments on various vision tasks demonstrate that our proposed framework can learn better generalized features for the unseen target domain compared with state of-the-art domain generalization methods.

Object detection typically assumes that training and test data are drawn from an identical distribution, which, however, does not always hold in practice. Such a distribution mismatch will lead to a significant performance drop. In this work, we aim to improve the cross-domain robustness of object detection. We tackle the domain shift on two levels: 1) the image-level shift, such as image style, illumination, etc, and 2) the instance-level shift, such as object appearance, size, etc. We build our approach based on the recent state-of-the-art Faster R-CNN model, and design two domain adaptation components, on image level and instance level, to reduce the domain discrepancy. The two domain adaptation components are based on H-divergence theory, and are implemented by learning a domain classifier in adversarial training manner. The domain classifiers on different levels are further reinforced with a consistency regularization to learn a domain-invariant region proposal network (RPN) in the Faster R-CNN model. We evaluate our newly proposed approach using multiple datasets including Cityscapes, KITTI, SIM10K, etc. The results demonstrate the effectiveness of our proposed approach for robust object detection in various domain shift scenarios.

This paper re-visits classical problems in document enhancement. Rather than proposing a new algorithm for a specific problem, we introduce a novel general approach. The key idea is to modify any state- of-the-art algorithm, by providing it with new information (input), improving its own results. Interestingly, this information is based on a solution to a seemingly unrelated problem of visibility detection in R3. We show that a simple representation of an image as a 3D point cloud, gives visibility detection on this cloud a new interpretation. What does it mean for a point to be visible? Although this question has been widely studied within computer vision, it has always been assumed that the point set is a sampling of a real scene. We show that the answer to this question in our context reveals unique and useful information about the image. We demonstrate the benefit of this idea for document binarization and for unshadowing.

Capturing document images is a common way for digitizing and recording physical documents due to the ubiquitousness of mobile cameras. To make text recognition easier, it is often desirable to digitally flatten a document image when the physical document sheet is folded or curved. In this paper, we develop the first learning-based method to achieve this goal. We propose a stacked U-Net with intermediate supervision to directly predict the forward mapping from a distorted image to its rectified version. Because large-scale real-world data with ground truth deformation is difficult to obtain, we create a synthetic dataset with approximately 100 thousand images by warping non-distorted document images. The network is trained on this dataset with various data augmentations to improve its generalization ability. We further create a comprehensive benchmark that covers various real-world conditions. We evaluate the proposed model quantitatively and qualitatively on the proposed benchmark, and compare it with previous non-learning-based methods.

Automated counting of people in crowd images is a challenging task. The major difficulty stems from the large diversity in the way people appear in crowds. In fact, features available for crowd discrimination largely depend on the crowd density to the extent that people are only seen as blobs in a highly dense scene. We tackle this problem with a growing CNN which can progressively increase its capacity to account for the wide variability seen in crowd scenes. Our model starts from a base CNN density regressor, which is trained in equivalence on all types of crowd images. In order to adapt with the huge diversity, we create two child regressors which are exact copies of the base CNN. A differential training procedure divides the dataset into two clusters and fine-tunes the child networks on their respective specialties. Consequently, without any hand-crafted criteria for forming specialties, the child regressors become experts on certain types of crowds. The child networks are again split recursively, creating two experts at every division. This hierarchical training leads to a CNN tree, where the child regressors are more fine experts than any of their parents. The leaf nodes are taken as the final experts and a classifier network is then trained to predict the correct specialty for a given test image patch. The proposed model achieves higher count accuracy on major crowd datasets. Further, we analyse the characteristics of specialties mined automatically by our method.

The increasing popularity of computational light field (LF) cameras has necessitated the need for tackling motion blur which is a ubiquitous phenomenon in hand-held photography. The state-of-the-art method for blind deblurring of LFs of general 3D scenes is limited to handling only downsampled LF, both in spatial and angular resolution. This is due to the computational overhead involved in processing data-hungry full-resolution 4D LF altogether. Moreover, the method warrants high-end GPUs for optimization and is ineffective for wide-angle settings and irregular camera motion. In this paper, we introduce a new blind motion deblurring strategy for LFs which alleviates these limitations significantly. Our model achieves this by isolating 4D LF motion blur across the 2D subaperture images, thus paving the way for independent deblurring of these subaperture images. Furthermore, our model accommodates common camera motion parameterization across the subaperture images. Consequently, blind deblurring of any single subaperture image elegantly paves the way for cost-effective non-blind deblurring of the other subaperture images. Our approach is CPU-efficient computationally and can effectively deblur full-resolution LFs.

Video-based person re-identification matches video clips of people across non-overlapping cameras. Most existing methods tackle this problem by encoding each video frame in its entirety and computing an aggregate representation across all frames. In practice, people are often partially occluded, which can corrupt the extracted features. Instead, we propose a new spatiotemporal attention model that automatically discovers a diverse set of distinctive body parts. This allows useful information to be extracted from all frames without succumbing to occlusions and misalignments. The network learns multiple spatial attention models and employs a diversity regularization term to ensure multiple models do not discover the same body part. Features extracted from local image regions are organized by spatial attention model and are combined using temporal attention. As a result, the network learns latent representations of the face, torso and other body parts using the best available image patches from the entire video sequence. Extensive evaluations on three datasets show that our framework outperforms the state-of-the-art approaches by large margins on multiple metrics.