ECCV 2020 论文列表

Photorealistic style transfer is the task of transferring the artistic style of an image onto a content target, producing a result that is plausibly taken with a camera. Recent approaches, based on deep neural networks, produce impressive results but are either too slow to run at practical resolutions, or still contain objectionable artifacts. We propose a new end-to-end model for photorealistic style transfer that is both fast and inherently generates photorealistic results. The core of our approach is a feed-forward neural network that learns local edge-aware affine transforms that automatically obey the photorealism constraint. When trained on a diverse set of images and a variety of styles, our model can robustly apply style transfer to an arbitrary pair of input images. Compared to the state of the art, our method produces visually superior results and is three orders of magnitude faster, enabling real-time performance at 4K on a mobile phone. We validate our method with ablation and user studies."

In this paper, we propose a method which jointly learns layout prediction and depth estimation from a single indoor panorama image. Previous methods have considered layout prediction and depth estimation from a single panorama image separately. However, these two tasks are tightly intertwined. Leveraging the layout depth map as an intermediate representation, our proposed method outperforms existing methods for both panorama layout prediction and depth estimation. Experiments on the challenging real-world dataset of Stanford 2D-3D demonstrate that our approach obtains superior performance for both the layout prediction tasks (3D IoU: 85.81% v.s. 79.79%) and the depth estimation (Abs Rel: 0.068 v.s. 0.079)."

Snow removal usually affects the performance of computer vision. Comparing with other atmospheric phenomenon (e.g., haze and rain), snow is more complicated due to its transparency, various size, and accumulation of veiling effect, which make single image de-snowing more challenging. In this paper, first, we reformulate the snow model. Different from that in the previous works, in the proposed snow model, the veiling effect is included. Second, a novel joint size and transparency-aware snow removal algorithm called JSTASR is proposed. It can classify snow particles according to their sizes and conduct snow removal in different scales. Moreover, to remove the snow with different transparency, the transparency-aware snow removal is developed. It can address both transparent and non-transparent snow particles by applying the modified partial convolution. Experiments show that the proposed method achieves significant improvement on both synthetic and real-world datasets and is very helpful for object detection on snow images."

Segmentation, and Registration System for 3D Multi-Modal Image Alignment of Large-scale Pathological CT Scans","Multi-modal image registration is a challenging problem that is also an important clinical task for many real applications and scenarios. As a first step in analysis, deformable registration among different image modalities is often required in order to provide complementary visual information. During registration, semantic information is key to match homologous points and pixels. Nevertheless, many conventional registration methods are incapable in capturing high-level semantic anatomical dense correspondences. In this work, we propose a novel multi-task learning system, JSSR, based on an end-to-end 3D convolutional neural network that is composed of a generator, a registration and a segmentation component. The system is optimized to satisfy the implicit constraints between different tasks in an unsupervised manner. It first synthesizes the source domain images into the target domain, then an intra-modal registration is applied on the synthesized images and target images. The segmentation module are then applied on the synthesized and target images, providing additional cues based on semantic correspondences. The supervision from another fully-annotated dataset is used to regularize the segmentation. We extensively evaluate JSSR on a large-scale medical image dataset containing 1,485 patient CT imaging studies of four different contrast phases (i.e., 5,940 3D CT scans with pathological livers) on the registration, segmentation and synthesis tasks. The performance is improved after joint training on the registration and segmentation tasks by 0.9% and 1.9% respectively compared to a highly competitive and accurate deep learning baseline. The registration also consistently outperforms conventional state-of-the-art multi-modal registration methods."

Semantic segmentation and semantic edge detection can be seen as two dual problems with close relationships in computer vision. Despite the fast evolution of learning-based 3D semantic segmentation methods, little attention has been drawn to the learning of 3D semantic edge detectors, even less to a joint learning method for the two tasks. In this paper, we tackle the 3D semantic edge detection task for the first time and present a new two-stream fully-convolutional network that jointly performs the two tasks. In particular, we design a joint refinement module that explicitly wires region information and edge information to improve the performances of both tasks. Further, we propose a novel loss function that encourages the network to produce semantic segmentation results with better boundaries. Extensive evaluations on S3DIS and ScanNet datasets show that our method achieves on par or better performance than the state-of-the-art methods for semantic segmentation and outperforms the baseline methods for semantic edge detection. "

In this paper, we introduce a novel approach to learn a 3D face model using a joint-based face rig and a neural skinning network. Thanks to the joint-based representation, our model enjoys some significant advantages over prior blendshape-based models. First, it is very compact such that we are orders of magnitude smaller while still keeping strong modeling capacity. Second, because each joint has its semantic meaning, interactive facial geometry editing is made easier and more intuitive. Third, through skinning, our model supports adding mouth interior and eyes, as well as accessories (hair, eye glasses, etc.) in a simpler, more accurate and principled way. We argue that because the human face is highly structured and topologically consistent, it does not need to be learned entirely from data. Instead we can leverage prior knowledge in the form of a human-designed 3D face rig to reduce the data dependency, and learn a compact yet strong face model from only a small dataset (less than one hundred 3D scans). To further improve the modeling capacity, we train a skinning weight generator through adversarial learning. Experiments on fitting high-quality 3D scans (both neutral and expressive), noisy depth images, and RGB images demonstrate that its modeling capacity is on-par with state-of-the-art face models, such as FLAME and Facewarehouse, even though the model is 10 to 20 times smaller. This suggests broad value in both graphics and vision applications on mobile and edge devices."

State-of-the-art single depth image-based 3D hand pose estimation methods are based on dense predictions, including voxel-to-voxel predictions, point-to-point regression, and pixel-wise estimations. Despite the good performance, those methods have a few issues in nature, such as the poor trade-off between accuracy and efficiency, and plain feature representation learning with local convolutions. In this paper, a novel pixel-wise prediction-based method is proposed to address the above issues. The key ideas are two-fold: a) explicitly modeling the dependencies among joints and the relations between the pixels and the joints for better local feature representation learning; b) unifying the dense pixel-wise offset predictions and direct joint regression for end-to-end training. Specifically, we first propose a graph convolutional network (GCN) based joint graph reasoning module to model the complex dependencies among joints and augment the representation capability of each pixel. Then we densely estimate all pixels' offsets to joints in both image plane and depth space and calculate the joints' positions by a weighted average over all pixels' predictions, totally discarding the complex post-processing operations. The proposed model is implemented with an efficient 2D fully convolutional network (FCN) backbone and has only about 1.4M parameters. Extensive experiments on multiple 3D hand pose estimation benchmarks demonstrate that the proposed method achieves new state-of-the-art accuracy while running very efficiently with around a speed of 110fps on a single NVIDIA 1080Ti GPU."

The general framework for fast universal style transfer consists of an autoencoder and a feature transformation at the bottleneck. We propose a new transformation that iteratively stylizes features with analytical gradient descent. Experiments show this transformation is advantageous in part because it is fast. With control knobs to balance content preservation and style effect transferal, we also show this method can switch between artistic and photo-realistic style transfers and reduce distortion and artifacts. Finally, we show it can be used for applications requiring spatial control and multiple-style transfer. "

In this paper, we propose a novel learning-based pipeline for partially overlapping 3D point cloud registration. The proposed model includes an iterative distance-aware similarity matrix convolution module to incorporate information from both the feature and Euclidean space into the pairwise point matching process. These convolution layers learn to match points based on joint information of the entire geometric features and Euclidean offset for each point pair, overcoming the disadvantage of matching by simply taking the inner product of feature vectors. Furthermore, a two-stage learnable point elimination technique is presented to improve computational efficiency and reduce false positive correspondence pairs. A novel mutual-supervision loss is proposed to train the model without extra annotations of keypoints. The pipeline can be easily integrated with both traditional (e.g. FPFH) and learning-based features. Experiments on partially overlapping and noisy point cloud registration show that our method outperforms the current state-of-the-art, while being more computationally efficient."

Human trajectory forecasting with multiple socially interact-ing agents is of critical importance for autonomous navigation in human environments, e.g., for self-driving cars and social robots. In this work, we present Predicted Endpoint Conditioned Network (PECNet) for flexible human trajectory prediction. PECNet infers distant trajectory endpoints to assist in long-range multi-modal trajectory prediction. A novel non-local social pooling layer enables PECNet to infer diverse yet socially compliant trajectories. Additionally, we present a simple “truncation-trick” for improving few-shot multi-modal trajectory prediction performance. We show that PECNet improves state-of-the-art performance on the Stanford Drone trajectory prediction benchmark by ~20.9% and on the ETH/UCY benchmark by ∼40.8%. Code available at the home page: https://karttikeya.github.io/publication/htf/"

Easy availability of wearable egocentric cameras, and the sense of privacy propagated by the fact that the wearer is never seen in the captured videos, has led to a tremendous rise in public sharing of such videos. Unlike hand-held cameras, egocentric cameras are harnessed on the wearer’s head, which makes it possible to track the wearer’s head motion by observing optical flow in the egocentric videos. In this work, we create a novel kind of privacy attack by extracting the wearer’s gait profile, a well known biometric signature, from such optical flow in the egocentric videos. We demonstrate strong wearer recognition capabilities based on extracted gait features, an unprecedented and critical weakness completely absent in hand-held videos. We demonstrate the following attack scenarios: (1) In a closed-set scenario, we show that it is possible to recognize the wearer of an egocentric video with an accuracy of more than 92.5% on the benchmark video dataset. (2) In an open-set setting, when the system has not seen the camera wearer even once during the training, we show that it is still possible to identify that the two egocentric videos have been captured by the same wearer with an Equal Error Rate (EER) of less than 14.35%. (3) We show that it is possible to extract gait signature even if only sparse optical flow and no other scene information from egocentric video is available. We demonstrate the accuracy of more than 84% for wearer recognition with only global optical flow. (4) While the first person to first person matching does not give us access to the wearer’s face, we show that it is possible to match the extracted gait features against the one obtained from a third person view such as a surveillance camera looking at the wearer in a completely different background at a different time. In essence, our work indicates that sharing one’s egocentric video should be treated as giving away one’s biometric identity and recommend much more oversight before sharing of egocentric videos. The code, trained models, and the datasets and their annotations are available at https://egocentricbiometric.github.io/"

Deep generative models have been successfully applied to Zero-Shot Learning (ZSL) recently. However, the underlying drawbacks of GANs and VAEs (e.g., the hardness of training with ZSL-oriented regularizers and the limited generation quality) hinder the existing generative ZSL models from fully bypassing the seen-unseen bias. To tackle the above limitations, for the first time, this work incorporates a new family of generative models (i.e., flow-based models) into ZSL. The proposed Invertible Zero-shot Flow (IZF) learns factorized data embeddings (i.e., the semantic factors and the non-semantic ones) with the forward pass of an invertible flow network, while the reverse pass generates data samples. This procedure theoretically extends conventional generative flows to a conditional scheme. To explicitly solve the bias problem, our model enlarges the seen-unseen distributional discrepancy based on negative sample-based distance measurement. Notably, IZF works flexibly with either a naive Bayesian classifier or a held-out trainable one for zero-shot recognition. Experiments on widely-adopted ZSL benchmarks demonstrate the significant performance gain of IZF over existing methods, in terms of both classic and generalized settings."

We introduce a novel neural network-based BRDF model and a Bayesian framework for object inverse rendering, i.e., joint estimation of reflectance and natural illumination from a single image of an object of known geometry. The BRDF is expressed with an invertible neural network, namely, normalizing flow, which provides the expressive power of a high-dimensional representation, computational simplicity of a compact analytical model, and physical plausibility of a real-world BRDF. We extract the latent space of real-world reflectance by conditioning this model, which directly results in a strong reflectance prior. We refer to this model as the invertible neural BRDF model (iBRDF). We also devise a deep illumination prior by leveraging the structural bias of deep neural networks. By integrating this novel BRDF model and reflectance and illumination priors in a MAP estimation formulation, we show that this joint estimation can be computed efficiently with stochastic gradient descent. We experimentally validate the accuracy of the invertible neural BRDF model on a large number of measured data and demonstrate its use in object inverse rendering on a number of synthetic and real images. The results show new ways in which deep neural networks can help solve challenging radiometric inverse problems."

High-resolution digital images are usually downscaled to fit various display screens or save the cost of storage and bandwidth, meanwhile the post-upscaling is adpoted to recover the original resolutions or the details in the zoom-in images. However, typical image downscaling is a non-injective mapping due to the loss of high-frequency information, which leads to the ill-posed problem of the inverse upscaling procedure and poses great challenges for recovering details from the downscaled low-resolution images. Simply upscaling with image super-resolution methods results in unsatisfactory recovering performance. In this work, we propose to solve this problem by modeling the downscaling and upscaling processes from a new perspective, i.e. an invertible bijective transformation, which can largely mitigate the ill-posed nature of image upscaling. We develop an Invertible Rescaling Net (IRN) with deliberately designed framework and objectives to produce visually-pleasing low-resolution images and meanwhile capture the distribution of the lost information using a latent variable following a specified distribution in the downscaling process. In this way, upscaling is made tractable by inversely passing a randomly-drawn latent variable with the low-resolution image through the network. Experimental results demonstrate the significant improvement of our model over existing methods in terms of both quantitative and qualitative evaluations of image upscaling reconstruction from downscaled images. Code is available at https://github.com/pkuxmq/Invertible-Image-Rescaling."

We present a learning-based method for interpolating and manipulating 3D shapes represented as point clouds, that is explicitly designed to preserve intrinsic shape properties. Our approach is based on constructing a dual encoding space that enables shape synthesis and, at the same time, provides links to the intrinsic shape metric, which is typically not available on point cloud data. Our method works in a single pass and avoids expensive optimization, employed by existing techniques. Furthermore, the strong regularization provided by our dual latent space approach also helps to improve shape recovery in challenging settings from noisy point clouds across datasets. Extensive experiments show that our method results in more realistic and smoother interpolations compared to baselines."

Current state-of-the-art semantic segmentation methods usually require high computational resources for accurate segmentation. One promising way to achieve a good trade-off between segmentation accuracy and efficiency is knowledge distillation. In this paper, different from previous methods performing knowledge distillation for densely pairwise relations, we propose a novel intra-class feature variation distillation (IFVD) to transfer the intra-class feature variation (IFV) of the cumbersome model (teacher) to the compact model (student). Concretely, we compute the feature center (regarded as the prototype) of each class and characterize the IFV with the set of similarity between the feature on each pixel and its corresponding class-wise prototype. The teacher model usually learns more robust intra-class feature representation than the student model, making them have different IFV. Transferring such IFV from teacher to student could make the student mimic the teacher better in terms of feature distribution, and thus improve the segmentation accuracy. We evaluate the proposed approach on three widely adopted benchmarks: Cityscapes, CamVid and Pascal VOC 2012, consistently improving state-of-the-art methods."

For person re-identification, existing deep networks often focus on representation learning. However, without transfer learning, the learned model is fixed as is, which is not adaptable for handling various unseen scenarios. In this paper, beyond representation learning, we consider how to formulate person image matching directly in deep feature maps. We treat image matching as finding local correspondences in feature maps, and construct query-adaptive convolution kernels on the fly to achieve local matching. In this way, the matching process and results are interpretable, and this explicit matching is more generalizable than representation features to unseen scenarios, such as unknown misalignments, pose or viewpoint changes. To facilitate end-to-end training of this architecture, we further build a class memory module to cache feature maps of the most recent samples of each class, so as to compute image matching losses for metric learning. Through direct cross-dataset evaluation, the proposed Query-Adaptive Convolution (QAConv) method gains large improvements over popular learning methods (about 10%+ mAP), and achieves comparable results to many transfer learning methods. Besides, a model-free temporal cooccurrence based score weighting method called TLift is proposed, which improves the performance to a further extent, achieving state-of-the-art results in cross-dataset person re-identification. Code is available at https://github.com/ShengcaiLiao/QAConv"

Visual reasoning is crucial for visual question answering (VQA). However, without labelled programs, implicit reasoning under natural supervision is still quite challenging and previous models are hard to interpret. In this paper, we rethink implicit reasoning process in VQA, and propose a new formulation which maximizes the log-likelihood of joint distribution for the observed question and predicted answer. Accordingly, we derive a Temporal Reasoning Network (TRN) framework which models the implicit reasoning process as sequential planning in latent space. Our model is interpretable on both model design in probabilist and reasoning process via visualization. We experimentally demonstrate that TRN can support implicit reasoning across various datasets. The experiment results of our model are competitive to existing implicit reasoning models and surpass baseline by large margin on complicated reasoning tasks without extra computation cost in forward stage."

The remarkable performance of convolutional neural networks (CNNs) is entangled with their huge number of uninterpretable parameters, which has become the bottleneck limiting the exploitation of their full potential. Towards network interpretation, previous endeavors mainly resort to the single filter analysis, which however ignores the relationship between filters. In this paper, we propose a novel architecture decoupling method to interpret the network from a perspective of investigating its calculation paths. More specifically, we introduce a novel architecture controlling module in each layer to encode the network architecture by a vector. By maximizing the mutual information between the vectors and input images, the module is trained to select specific filters to distill a unique calculation path for each input. Furthermore, to improve the interpretability and compactness of the decoupled network, the output of each layer is encoded to align the architecture encoding vector with the constraint of sparsity regularization. Unlike conventional pixel-level or filter-level network interpretation methods, we propose a path-level analysis to explore the relationship between the combination of filter and semantic concepts, which is more suitable to interpret the working rationale of the network. Extensive experiments show that the decoupled network achieves several applications, i.e., network interpretation, network acceleration, and adversarial samples detection."