ICCV 2019 论文列表

Video compression artifact reduction aims to recover high-quality videos from low-quality compressed videos. Most existing approaches use a single neighboring frame or a pair of neighboring frames (preceding and/or following the target frame) for this task. Furthermore, as frames of high quality overall may contain low-quality patches, and high-quality patches may exist in frames of low quality overall, current methods focusing on nearby peak-quality frames (PQFs) may miss high-quality details in low-quality frames. To remedy these shortcomings, in this paper we propose a novel end-to-end deep neural network called non-local ConvLSTM (NL-ConvLSTM in short) that exploits multiple consecutive frames. An approximate non-local strategy is introduced in NL-ConvLSTM to capture global motion patterns and trace the spatiotemporal dependency in a video sequence. This approximate strategy makes the non-local module work in a fast and low space-cost way. Our method uses the preceding and following frames of the target frame to generate a residual, from which a higher quality frame is reconstructed. Experiments on two datasets show that NL-ConvLSTM outperforms the existing methods.

We show that for human-object interaction detection a relatively simple factorized model with appearance and layout encodings constructed from pre-trained object detectors outperforms more sophisticated approaches. Our model includes factors for detection scores, human and object appearance, and coarse (box-pair configuration) and optionally fine-grained layout (human pose). We also develop training techniques that improve learning efficiency by: (1) eliminating a train-inference mismatch; (2) rejecting easy negatives during mini-batch training; and (3) using a ratio of negatives to positives that is two orders of magnitude larger than existing approaches. We conduct a thorough ablation study to understand the importance of different factors and training techniques using the challenging HICO-Det dataset.

Image retrieval under varying illumination conditions, such as day and night images, is addressed by image preprocessing, both hand-crafted and learned. Prior to extracting image descriptors by a convolutional neural network, images are photometrically normalised in order to reduce the descriptor sensitivity to illumination changes. We propose a learnable normalisation based on the U-Net architecture, which is trained on a combination of single-camera multi-exposure images and a newly constructed collection of similar views of landmarks during day and night. We experimentally show that both hand-crafted normalisation based on local histogram equalisation and the learnable normalisation outperform standard approaches in varying illumination conditions, while staying on par with the state-of-the-art methods on daylight illumination benchmarks, such as Oxford or Paris datasets.

Treating graph structures of Markov random fields as unknown and estimating them jointly with labels have been shown to be useful for modeling human activity recognition and other related tasks. We propose two novel relaxations for solving this problem. The first is a linear programming (LP) relaxation, which is provably tighter than the existing LP relaxation. The second is a non-convex quadratic programming (QP) relaxation, which admits an efficient concave-convex procedure (CCCP). The CCCP algorithm is initialized by solving a convex QP relaxation of the problem, which is obtained by modifying the diagonal of the matrix that specifies the non-convex QP relaxation. We show that our convex QP relaxation is optimal in the sense that it minimizes the L1 norm of the diagonal modification vector. While the convex QP relaxation is not as tight as the existing and the new LP relaxations, when used in conjunction with the CCCP algorithm for the non-convex QP relaxation, it provides accurate solutions. We demonstrate the efficacy of our new relaxations for both synthetic data and human activity recognition.

We present Neural-Guided RANSAC (NG-RANSAC), an extension to the classic RANSAC algorithm from robust optimization. NG-RANSAC uses prior information to improve model hypothesis search, increasing the chance of finding outlier-free minimal sets. Previous works use heuristic side-information like hand-crafted descriptor distance to guide hypothesis search. In contrast, we learn hypothesis search in a principled fashion that lets us optimize an arbitrary task loss during training, leading to large improvements on classic computer vision tasks. We present two further extensions to NG-RANSAC. Firstly, using the inlier count itself as training signal allows us to train neural guidance in a self-supervised fashion. Secondly, we combine neural guidance with differentiable RANSAC to build neural networks which focus on certain parts of the input data and make the output predictions as good as possible. We evaluate NG-RANSAC on a wide array of computer vision tasks, namely estimation of epipolar geometry, horizon line estimation and camera re-localization. We achieve superior or competitive results compared to state-of-the-art robust estimators, including very recent, learned ones.

We introduce a method to generate videos of dynamic virtual objects plausibly interacting via collisions with a still image's environment. Given a starting trajectory, physically simulated with the estimated geometry of a single, static input image, we learn to 'correct' this trajectory to a visually plausible one via a neural network. The neural network can then be seen as learning to 'correct' traditional simulation output, generated with incomplete and imprecise world information, to obtain context-specific, visually plausible re-simulated output - a process we call neural re-simulation. We train our system on a set of 50k synthetic scenes where a virtual moving object (ball) has been physically simulated. We demonstrate our approach on both our synthetic dataset and a collection of real-life images depicting everyday scenes, obtaining consistent improvement over baseline alternatives throughout.

Inverse rendering aims to estimate physical attributes of a scene, e.g., reflectance, geometry, and lighting, from image(s). Inverse rendering has been studied primarily for single objects or with methods that solve for only one of the scene attributes. We propose the first learning based approach that jointly estimates albedo, normals, and lighting of an indoor scene from a single image. Our key contribution is the Residual Appearance Renderer (RAR), which can be trained to synthesize complex appearance effects (e.g., inter-reflection, cast shadows, near-field illumination, and realistic shading), which would be neglected otherwise. This enables us to perform self-supervised learning on real data using a reconstruction loss, based on re-synthesizing the input image from the estimated components. We finetune with real data after pretraining with synthetic data. To this end, we use physically-based rendering to create a large-scale synthetic dataset, named SUNCG-PBR, which is a significant improvement over prior datasets. Experimental results show that our approach outperforms state-of-the-art methods that estimate one or more scene attributes.

While there are many deep learning based approaches for single image compression, the field of end-to-end learned video coding has remained much less explored. Therefore, in this work we present an inter-frame compression approach for neural video coding that can seamlessly build up on different existing neural image codecs. Our end-to-end solution performs temporal prediction by optical flow based motion compensation in pixel space. The key insight is that we can increase both decoding efficiency and reconstruction quality by encoding the required information into a latent representation that directly decodes into motion and blending coefficients. In order to account for remaining prediction errors, residual information between the original image and the interpolated frame is needed. We propose to compute residuals directly in latent space instead of in pixel space as this allows to reuse the same image compression network for both key frames and intermediate frames. Our extended evaluation on different datasets and resolutions shows that the rate-distortion performance of our approach is competitive with existing state-of-the-art codecs.

Generative models for 3D geometric data arise in many important applications in 3D computer vision and graphics. In this paper, we focus on 3D deformable shapes that share a common topological structure, such as human faces and bodies. Morphable Models and their variants, despite their linear formulation, have been widely used for shape representation, while most of the recently proposed nonlinear approaches resort to intermediate representations, such as 3D voxel grids or 2D views. In this work, we introduce a novel graph convolutional operator, acting directly on the 3D mesh, that explicitly models the inductive bias of the fixed underlying graph. This is achieved by enforcing consistent local orderings of the vertices of the graph, through the spiral operator, thus breaking the permutation invariance property that is adopted by all the prior work on Graph Neural Networks. Our operator comes by construction with desirable properties (anisotropic, topology-aware, lightweight, easy-to-optimise), and by using it as a building block for traditional deep generative architectures, we demonstrate state-of-the-art results on a variety of 3D shape datasets compared to the linear Morphable Model and other graph convolutional operators.

In this paper, we propose a Neighborhood Preserving Hashing (NPH) method for scalable video retrieval in an unsupervised manner. Unlike most existing deep video hashing methods which indiscriminately compress an entire video into a binary code, we embed the spatial-temporal neighborhood information into the encoding network such that the neighborhood-relevant visual content of a video can be preferentially encoded into a binary code under the guidance of the neighborhood information. Specifically, we propose a neighborhood attention mechanism which focuses on partial useful content of each input frame conditioned on the neighborhood information. We then integrate the neighborhood attention mechanism into an RNN-based reconstruction scheme to encourage the binary codes to capture the spatial-temporal structure in a video which is consistent with that in the neighborhood. As a consequence, the learned hashing functions can map similar videos to similar binary codes. Extensive experiments on three widely-used benchmark datasets validate the effectiveness of our proposed approach.

The labeling cost of large number of bounding boxes is one of the main challenges for training modern object detectors. To reduce the dependence on expensive bounding box annotations, we propose a new semi-supervised object detection formulation, in which a few seed box level annotations and a large scale of image level annotations are used to train the detector. We adopt a training-mining framework, which is widely used in weakly supervised object detection tasks. However, the mining process inherently introduces various kinds of labelling noises: false negatives, false positives and inaccurate boundaries, which can be harmful for training the standard object detectors (e.g. Faster RCNN). We propose a novel NOise Tolerant Ensemble RCNN (NOTE-RCNN) object detector to handle such noisy labels. Comparing to standard Faster RCNN, it contains three highlights: an ensemble of two classification heads and a distillation head to avoid overfitting on noisy labels and improve the mining precision, masking the negative sample loss in box predictor to avoid the harm of false negative labels, and training box regression head only on seed annotations to eliminate the harm from inaccurate boundaries of mined bounding boxes. We evaluate the methods on ILSVRC 2013 and MSCOCO 2017 dataset; we observe that the detection accuracy consistently improves as we iterate between mining and training steps, and state-of-the-art performance is achieved.

Convolutional Neural Networks (CNNs) provide excellent performance when used for image classification. The classical method of training CNNs is by labeling images in a supervised manner as in "input image belongs to this label" (Positive Learning; PL), which is a fast and accurate method if the labels are assigned correctly to all images. However, if inaccurate labels, or noisy labels, exist, training with PL will provide wrong information, thus severely degrading performance. To address this issue, we start with an indirect learning method called Negative Learning (NL), in which the CNNs are trained using a complementary label as in "input image does not belong to this complementary label." Because the chances of selecting a true label as a complementary label are low, NL decreases the risk of providing incorrect information. Furthermore, to improve convergence, we extend our method by adopting PL selectively, termed as Selective Negative Learning and Positive Learning (SelNLPL). PL is used selectively to train upon expected-to-be-clean data, whose choices become possible as NL progresses, thus resulting in superior performance of filtering out noisy data. With simple semi-supervised training technique, our method achieves state-of-the-art accuracy for noisy data classification, proving the superiority of SelNLPL's noisy data filtering ability.

An assumption widely used in recent neural style transfer methods is that image styles can be described by global statics of deep features like Gram or covariance matrices. Alternative approaches have represented styles by decomposing them into local pixel or neural patches. Despite the recent progress, most existing methods treat the semantic patterns of style image uniformly, resulting unpleasing results on complex styles. In this paper, we introduce a more flexible and general universal style transfer technique: multimodal style transfer (MST). MST explicitly considers the matching of semantic patterns in content and style images. Specifically, the style image features are clustered into sub-style components, which are matched with local content features under a graph cut formulation. A reconstruction network is trained to transfer each sub-style and render the final stylized result. We also generalize MST to improve some existing methods. Extensive experiments demonstrate the superior effectiveness, robustness, and flexibility of MST.

Existing deep learning-based interactive image segmentation approaches typically assume the target-of-interest is always a single object and fail to account for the potential diversity in user expectations, thus requiring excessive user input when it comes to segmenting an object part or a group of objects instead. Motivated by the observation that the object part, full object, and a collection of objects essentially differ in size, we propose a new concept called scale-diversity, which characterizes the spectrum of segmentations w.r.t. different scales. To address this, we present MultiSeg, a scale-diverse interactive image segmentation network that incorporates a set of two-dimensional scale priors into the model to generate a set of scale-varying proposals that conform to the user input. We explicitly encourage segmentation diversity during training by synthesizing diverse training samples for a given image. As a result, our method allows the user to quickly locate the closest segmentation target for further refinement if necessary. Despite its simplicity, experimental results demonstrate that our proposed model is capable of quickly producing diverse yet plausible segmentation outputs, reducing the user interaction required, especially in cases where many types of segmentations (object parts or groups) are expected.

We propose a novel idea for depth estimation from multi-view image-pose pairs, where the model has capability to leverage information from previous latent-space encodings of the scene. This model uses pairs of images and poses, which are passed through an encoder-decoder model for disparity estimation. The novelty lies in soft-constraining the bottleneck layer by a nonparametric Gaussian process prior. We propose a pose-kernel structure that encourages similar poses to have resembling latent spaces. The flexibility of the Gaussian process (GP) prior provides adapting memory for fusing information from nearby views. We train the encoder-decoder and the GP hyperparameters jointly end-to-end. In addition to a batch method, we derive a lightweight estimation scheme that circumvents standard pitfalls in scaling Gaussian process inference, and demonstrate how our scheme can run in real-time on smart devices.

We present an end-to-end learned system for fusing multiple misaligned photographs of the same scene into a chosen target view. We demonstrate three use cases: 1) color transfer for inferring color for a monochrome view, 2) HDR fusion for merging misaligned bracketed exposures, and 3) detail transfer for reprojecting a high definition image to the point of view of an affordable VR180-camera. While the system can be trained end-to-end, it consists of three distinct steps: feature extraction, image warping and fusion. We present a novel cascaded feature extraction method that enables us to synergetically learn optical flow at different resolution levels. We show that this significantly improves the network's ability to learn large disparities. Finally, we demonstrate that our alignment architecture outperforms a state-of-the art optical flow network on the image warping task when both systems are trained in an identical manner.

Exploiting relationships between visual regions and question words have achieved great success in learning multi-modality features for Visual Question Answering (VQA). However, we argue that existing methods mostly model relations between individual visual regions and words, which are not enough to correctly answer the question. From humans' perspective, answering a visual question requires understanding the summarizations of visual and language information. In this paper, we proposed the Multi-modality Latent Interaction module (MLI) to tackle this problem. The proposed module learns the cross-modality relationships between latent visual and language summarizations, which summarize visual regions and question into a small number of latent representations to avoid modeling uninformative individual region-word relations. The cross-modality information between the latent summarizations are propagated to fuse valuable information from both modalities and are used to update the visual and word features. Such MLI modules can be stacked for several stages to model complex and latent relations between the two modalities and achieves highly competitive performance on public VQA benchmarks, VQA v2.0 and TDIUC . In addition, we show that the performance of our methods could be significantly improved by combining with pre-trained language model BERT.

Crowd counting presents enormous challenges in the form of large variation in scales within images and across the dataset. These issues are further exacerbated in highly congested scenes. Approaches based on straightforward fusion of multi-scale features from a deep network seem to be obvious solutions to this problem. However, these fusion approaches do not yield significant improvements in the case of crowd counting in congested scenes. This is usually due to their limited abilities in effectively combining the multi-scale features for problems like crowd counting. To overcome this, we focus on how to efficiently leverage information present in different layers of the network. Specifically, we present a network that involves: (i) a multi-level bottom-top and top-bottom fusion (MBTTBF) method to combine information from shallower to deeper layers and vice versa at multiple levels, (ii) scale complementary feature extraction blocks (SCFB) involving cross-scale residual functions to explicitly enable flow of complementary features from adjacent conv layers along the fusion paths. Furthermore, in order to increase the effectiveness of the multi-scale fusion, we employ a principled way of generating scale-aware ground-truth density maps for training. Experiments conducted on three datasets that contain highly congested scenes (ShanghaiTech, UCF_CC_50, and UCF-QNRF) demonstrate that the proposed method is able to outperform several recent methods in all the datasets

We present Multi-Garment Network (MGN), a method to predict body shape and clothing, layered on top of the SMPL model from a few frames (1-8) of a video. Several experiments demonstrate that this representation allows higher level of control when compared to single mesh or voxel representations of shape. Our model allows to predict garment geometry, relate it to the body shape, and transfer it to new body shapes and poses. To train MGN, we leverage a digital wardrobe containing 712 digital garments in correspondence, obtained with a novel method to register a set of clothing templates to a dataset of real 3D scans of people in different clothing and poses. Garments from the digital wardrobe, or predicted by MGN, can be used to dress any body shape in arbitrary poses. We will make publicly available the digital wardrobe, the MGN model, and code to dress SMPL with the garments at https://virtualhumans.mpi-inf.mpg.de/mgn