ICCV 2019 Papers

We propose a method for unsupervised domain adaptation that trains a shared embedding to align the joint distributions of inputs (domain) and outputs (classes), making any classifier agnostic to the domain. Joint alignment ensures that not only the marginal distributions of the domains are aligned, but the labels as well. We propose a novel objective function that encourages the class-conditional distributions to have disjoint support in feature space. We further exploit adversarial regularization to improve the performance of the classifier on the domain for which no annotated data is available.

We present a novel method for computing correspondences across 3D shapes using unsupervised learning. Our method computes a non-linear transformation of given descriptor functions, while optimizing for global structural properties of the resulting maps, such as their bijectivity or approximate isometry. To this end, we use the functional maps framework, and build upon the recent FMNet architecture for descriptor learning. Unlike that approach, however, we show that learning can be done in a purely unsupervised setting, without having access to any ground truth correspondences. This results in a very general shape matching method that we call SURFMNet for Spectral Unsupervised FMNet, and which can be used to establish correspondences within 3D shape collections without any prior information. We demonstrate on a wide range of challenging benchmarks, that our approach leads to state-of-the-art results compared to the existing unsupervised methods and achieves results that are comparable even to the supervised learning techniques. Moreover, our framework is an order of magnitude faster, and does not rely on geodesic distance computation or expensive post-processing.

Deep learning models have shown their vulnerabilities to universal adversarial perturbations (UAP), which are quasi-imperceptible. Compared to the conventional supervised UAPs that suffer from the knowledge of training data, the data-independent unsupervised UAPs are more applicable. Existing unsupervised methods fail to take advantage of the model uncertainty to produce robust perturbations. In this paper, we propose a new unsupervised universal adversarial perturbation method, termed as Prior Driven Uncertainty Approximation (PD-UA), to generate a robust UAP by fully exploiting the model uncertainty at each network layer. Specifically, a Monte Carlo sampling method is deployed to activate more neurons to increase the model uncertainty for a better adversarial perturbation. Thereafter, a textural bias prior to revealing a statistical uncertainty is proposed, which helps to improve the attacking performance. The UAP is crafted by the stochastic gradient descent algorithm with a boosted momentum optimizer, and a Laplacian pyramid frequency model is finally used to maintain the statistical uncertainty. Extensive experiments demonstrate that our method achieves well attacking performances on the ImageNet validation set, and significantly improves the fooling rate compared with the state-of-the-art methods.

Style transfer is a task of rendering images in the styles of other images. In the past few years, neural style transfer has achieved a great success in this task, yet suffers from either the inability to generalize to unseen style images or fast style transfer. Recently, an universal style transfer technique that applies zero-phase component analysis (ZCA) for whitening and coloring image features realizes fast and arbitrary style transfer. However, using ZCA for style transfer is empirical and does not have any theoretical support. In addition, other whitening and coloring transforms (WCT) than ZCA have not been investigated. In this report, we generalize ZCA to the general form of WCT, provide an analytical performance analysis from the angle of neural style transfer, and show why ZCA is a good choice for style transfer among different WCTs and why some WCTs are not well applicable for style transfer.

Attribution is the problem of finding which parts of an image are the most responsible for the output of a deep neural network. An important family of attribution methods is based on measuring the effect of perturbations applied to the input image, either via exhaustive search or by finding representative perturbations via optimization. In this paper, we discuss some of the shortcomings of existing approaches to perturbation analysis and address them by introducing the concept of extremal perturbations, which are theoretically grounded and interpretable. We also introduce a number of technical innovations to compute these extremal perturbations, including a new area constraint and a parametric family of smooth perturbations, which allow us to remove all tunable weighing factors from the optimization problem. We analyze the effect of perturbations as a function of their area, demonstrating excellent sensitivity to the spatial properties of the network under stimulation. We also extend perturbation analysis to the intermediate layers of a deep neural network. This application allows us to show how compactly an image can be represented (in terms of the number of channels it requires). We also demonstrate that the consistency with which images of a given class rely on the same intermediate channel correlates well with class accuracy.

Recently, there has been a renewed interest in leveraging multiple cameras, but under unconstrained settings. They have been quite successfully deployed in smartphones, which have become de facto choice for many photographic applications. However, akin to normal cameras, the functionality of multi-camera systems can be marred by motion blur which is a ubiquitous phenomenon in hand-held cameras. Despite the far-reaching potential of unconstrained camera arrays, there is not a single deblurring method for such systems. In this paper, we propose a generalized blur model that elegantly explains the intrinsically coupled image formation model for dual-lens set-up, which are by far most predominant in smartphones. While image aesthetics is the main objective in normal camera deblurring, any method conceived for our problem is additionally tasked with ascertaining consistent scene-depth in the deblurred images. We reveal an intriguing challenge that stems from an inherent ambiguity unique to this problem which naturally disrupts this coherence. We address this issue by devising a judicious prior, and based on our model and prior propose a practical blind deblurring method for dual-lens cameras, that achieves state-of-the-art performance.

Deep neural networks (DNNs) provide state-of-the-art results for a multitude of applications, but the approaches using DNNs for multimodal audiovisual applications do not consider predictive uncertainty associated with individual modalities. Bayesian deep learning methods provide principled confidence and quantify predictive uncertainty. Our contribution in this work is to propose an uncertainty aware multimodal Bayesian fusion framework for activity recognition. We demonstrate a novel approach that combines deterministic and variational layers to scale Bayesian DNNs to deeper architectures. Our experiments using in- and out-of-distribution samples selected from a subset of Moments-in-Time (MiT) dataset show a more reliable confidence measure as compared to the non-Bayesian baseline and the Monte Carlo dropout (MC dropout) approximate Bayesian inference. We also demonstrate the uncertainty estimates obtained from the proposed framework can identify out-of-distribution data on the UCF101 and MiT datasets. In the multimodal setting, the proposed framework improved precision-recall AUC by 10.2% on the subset of MiT dataset as compared to non-Bayesian baseline.

Aiming towards human-level generalization, there is a need to explore adaptable representation learning methods with greater transferability. Most existing approaches independently address task-transferability and cross-domain adaptation, resulting in limited generalization. In this paper, we propose UM-Adapt - a unified framework to effectively perform unsupervised domain adaptation for spatially-structured prediction tasks, simultaneously maintaining a balanced performance across individual tasks in a multi-task setting. To realize this, we propose two novel regularization strategies; a) Contour-based content regularization (CCR) and b) exploitation of inter-task coherency using a cross-task distillation module. Furthermore, avoiding a conventional ad-hoc domain discriminator, we re-utilize the cross-task distillation loss as output of an energy function to adversarially minimize the input domain discrepancy. Through extensive experiments, we demonstrate superior generalizability of the learned representations simultaneously for multiple tasks under domain-shifts from synthetic to natural environments. UM-Adapt yields state-of-the-art transfer learning results on ImageNet classification and comparable performance on PASCAL VOC 2007 detection task, even with a smaller backbone-net. Moreover, the resulting semi-supervised framework outperforms the current fully-supervised multi-task learning state-of-the-art on both NYUD and Cityscapes dataset.

We propose a novel approach to performing fine-grained 3D manipulation of image content via a convolutional neural network, which we call the Transformable Bottleneck Network (TBN). It applies given spatial transformations directly to a volumetric bottleneck within our encoder-bottleneck-decoder architecture. Multi-view supervision encourages the network to learn to spatially disentangle the feature space within the bottleneck. The resulting spatial structure can be manipulated with arbitrary spatial transformations. We demonstrate the efficacy of TBNs for novel view synthesis, achieving state-of-the-art results on a challenging benchmark. We demonstrate that the bottlenecks produced by networks trained for this task contain meaningful spatial structure that allows us to intuitively perform a variety of image manipulations in 3D, well beyond the rigid transformations seen during training. These manipulations include non-uniform scaling, non-rigid warping, and combining content from different images. Finally, we extract explicit 3D structure from the bottleneck, performing impressive 3D reconstruction from a single input image.

We propose a novel approach for estimating the difficulty and transferability of supervised classification tasks. Unlike previous work, our approach is solution agnostic and does not require or assume trained models. Instead, we estimate these values using an information theoretic approach: treating training labels as random variables and exploring their statistics. When transferring from a source to a target task, we consider the conditional entropy between two such variables (i.e., label assignments of the two tasks). We show analytically and empirically that this value is related to the loss of the transferred model. We further show how to use this value to estimate task hardness. We test our claims extensively on three large scale data sets---CelebA (40 tasks), Animals with Attributes 2 (85 tasks), and Caltech-UCSD Birds 200 (312 tasks)---together representing 437 classification tasks. We provide results showing that our hardness and transferability estimates are strongly correlated with empirical hardness and transferability. As a case study, we transfer a learned face recognition model to CelebA attribute classification tasks, showing state of the art accuracy for highly transferable attributes.

Zero-shot object detection (ZSD) is a relatively unexplored research problem as compared to the conventional zero-shot recognition task. ZSD aims to detect previously unseen objects during inference. Existing ZSD works suffer from two critical issues: (a) large domain-shift between the source (seen) and target (unseen) domains since the two distributions are highly mismatched. (b) the learned model is biased against unseen classes, therefore in generalized ZSD settings, where both seen and unseen objects co-occur during inference, the learned model tends to misclassify unseen to seen categories. This brings up an important question: How effectively can a transductive setting address the aforementioned problems? To the best of our knowledge, we are the first to propose a transductive zero-shot object detection approach that convincingly reduces the domain-shift and model-bias against unseen classes. Our approach is based on a self-learning mechanism that uses a novel hybrid pseudo-labeling technique. It progressively updates learned model parameters by associating unlabeled data samples to their corresponding classes. During this process, our technique makes sure that knowledge that was previously acquired on the source domain is not forgotten. We report significant 'relative' improvements of 34.9% and 77.1% in terms of mAP and recall rates over the previous best inductive models on MSCOCO dataset.

We propose an end-to-end trainable network that can simultaneously detect and recognize text of arbitrary shape, making substantial progress on the open problem of reading scene text of irregular shape. We formulate arbitrary shape text detection as an instance segmentation problem; an attention model is then used to decode the textual content of each irregularly shaped text region without rectification. To extract useful irregularly shaped text instance features from image scale features, we propose a simple yet effective RoI masking step. Additionally, we show that predictions from an existing multi-step OCR engine can be leveraged as partially labeled training data, which leads to significant improvements in both the detection and recognition accuracy of our model. Our method surpasses the state-of-the-art for end-to-end recognition tasks on the ICDAR15 (straight) benchmark by 4.6%, and on the Total-Text (curved) benchmark by more than 16%.

Recent advancements in fields like Internet of Things (IoT), augmented reality, etc. have led to an unprecedented demand for miniature cameras with low cost that can be integrated anywhere and can be used for distributed monitoring. Mask-based lensless imaging systems make such inexpensive and compact models realizable. However, reduction in the size and cost of these imagers comes at the expense of their image quality due to the high degree of multiplexing inherent in their design. In this paper, we present a method to obtain image reconstructions from mask-based lensless measurements that are more photorealistic than those currently available in the literature. We particularly focus on FlatCam, a lensless imager consisting of a coded mask placed over a bare CMOS sensor. Existing techniques for reconstructing FlatCam measurements suffer from several drawbacks including lower resolution and dynamic range than lens-based cameras. Our approach overcomes these drawbacks using a fully trainable non-iterative deep learning based model. Our approach is based on two stages: an inversion stage that maps the measurement into the space of intermediate reconstruction and a perceptual enhancement stage that improves this intermediate reconstruction based on perceptual and signal distortion metrics. Our proposed method is fast and produces photo-realistic reconstruction as demonstrated on many real and challenging scenes.

This paper first proposes a method of formulating model interpretability in visual understanding tasks based on the idea of unfolding latent structures. It then presents a case study in object detection using popular two-stage region-based convolutional network (i.e., R-CNN) detection systems. The proposed method focuses on weakly-supervised extractive rationale generation, that is learning to unfold latent discriminative part configurations of object instances automatically and simultaneously in detection without using any supervision for part configurations. It utilizes a top-down hierarchical and compositional grammar model embedded in a directed acyclic AND-OR Graph (AOG) to explore and unfold the space of latent part configurations of regions of interest (RoIs). It presents an AOGParsing operator that seamlessly integrates with the RoIPooling/RoIAlign operator widely used in R-CNN and is trained end-to-end. In object detection, a bounding box is interpreted by the best parse tree derived from the AOG on-the-fly, which is treated as the qualitatively extractive rationale generated for interpreting detection. In experiments, Faster R-CNN is used to test the proposed method on the PASCAL VOC 2007 and the COCO 2017 object detection datasets. The experimental results show that the proposed method can compute promising latent structures without hurting the performance. The code and pretrained models are available at https://github.com/iVMCL/iRCNN.

Deep CNNs have been pushing the frontier of visual recognition over past years. Besides recognition accuracy, strong demands in understanding deep CNNs in the research community motivate developments of tools to dissect pre-trained models to visualize how they make predictions. Recent works further push the interpretability in the network learning stage to learn more meaningful representations. In this work, focusing on a specific area of visual recognition, we report our efforts towards interpretable face recognition. We propose a spatial activation diversity loss to learn more structured face representations. By leveraging the structure, we further design a feature activation diversity loss to push the interpretable representations to be discriminative and robust to occlusions. We demonstrate on three face recognition benchmarks that our proposed method is able to achieve the state-of-art face recognition accuracy with easily interpretable face representations.

Most of the existing learning-based single image super-resolution (SISR) methods are trained and evaluated on simulated datasets, where the low-resolution (LR) images are generated by applying a simple and uniform degradation (i.e., bicubic downsampling) to their high-resolution (HR) counterparts. However, the degradations in real-world LR images are far more complicated. As a consequence, the SISR models trained on simulated data become less effective when applied to practical scenarios. In this paper, we build a real-world super-resolution (RealSR) dataset where paired LR-HR images on the same scene are captured by adjusting the focal length of a digital camera. An image registration algorithm is developed to progressively align the image pairs at different resolutions. Considering that the degradation kernels are naturally non-uniform in our dataset, we present a Laplacian pyramid based kernel prediction network (LP-KPN), which efficiently learns per-pixel kernels to recover the HR image. Our extensive experiments demonstrate that SISR models trained on our RealSR dataset deliver better visual quality with sharper edges and finer textures on real-world scenes than those trained on simulated datasets. Though our RealSR dataset is built by using only two cameras (Canon 5D3 and Nikon D810), the trained model generalizes well to other camera devices such as Sony a7II and mobile phones.

In recent years, substantial progress has been achieved in learning-based reconstruction of 3D objects. At the same time, generative models were proposed that can generate highly realistic images. However, despite this success in these closely related tasks, texture reconstruction of 3D objects has received little attention from the research community and state-of-the-art methods are either limited to comparably low resolution or constrained experimental setups. A major reason for these limitations is that common representations of texture are inefficient or hard to interface for modern deep learning techniques. In this paper, we propose Texture Fields, a novel texture representation which is based on regressing a continuous 3D function parameterized with a neural network. Our approach circumvents limiting factors like shape discretization and parameterization, as the proposed texture representation is independent of the shape representation of the 3D object. We show that Texture Fields are able to represent high frequency texture and naturally blend with modern deep learning techniques. Experimentally, we find that Texture Fields compare favorably to state-of-the-art methods for conditional texture reconstruction of 3D objects and enable learning of probabilistic generative models for texturing unseen 3D models. We believe that Texture Fields will become an important building block for the next generation of generative 3D models.

Although various image-based domain adaptation (DA) techniques have been proposed in recent years, domain shift in videos is still not well-explored. Most previous works only evaluate performance on small-scale datasets which are saturated. Therefore, we first propose two large-scale video DA datasets with much larger domain discrepancy: UCF-HMDB_full and Kinetics-Gameplay. Second, we investigate different DA integration methods for videos, and show that simultaneously aligning and learning temporal dynamics achieves effective alignment even without sophisticated DA methods. Finally, we propose Temporal Attentive Adversarial Adaptation Network (TA3N), which explicitly attends to the temporal dynamics using domain discrepancy for more effective domain alignment, achieving state-of-the-art performance on four video DA datasets (e.g. 7.9% accuracy gain over "Source only" from 73.9% to 81.8% on "HMDB --> UCF", and 10.3% gain on "Kinetics --> Gameplay"). The code and data are released at http://github.com/cmhungsteve/TA3N.

Human pose and shape are two important components of 2D human body. However, how to efficiently represent both of them in images is still an open question. In this paper, we propose the Triplet Representation for Body (TRB) --- a compact 2D human body representation, with skeleton keypoints capturing human pose information and contour keypoints containing human shape information. TRB not only preserves the flexibility of skeleton keypoint representation, but also contains rich pose and human shape information. Therefore, it promises broader application areas, such as human shape editing and conditional image generation. We further introduce the challenging problem of TRB estimation, where joint learning of human pose and shape is required. We construct several large-scale TRB estimation datasets, based on the popular 2D pose datasets LSP, MPII and COCO. To effectively solve TRB estimation, we propose a two-branch network (TRB-net) with three novel techniques, namely X-structure (Xs), Directional Convolution (DC) and Pairwise mapping (PM), to enforce multi-level message passing for joint feature learning. We evaluate our proposed TRB-net and several leading approaches on our proposed TRB datasets, and demonstrate the superiority of our method through extensive evaluations.