ICCV 2019 Papers

Despite great progress in 3D pose estimation from single-view images or videos, it remains a challenging task due to the substantial depth ambiguity and severe self-occlusions. Motivated by the effectiveness of incorporating spatial dependencies and temporal consistencies to alleviate these issues, we propose a novel graph-based method to tackle the problem of 3D human body and 3D hand pose estimation from a short sequence of 2D joint detections. Particularly, domain knowledge about the human hand (body) configurations is explicitly incorporated into the graph convolutional operations to meet the specific demand of the 3D pose estimation. Furthermore, we introduce a local-to-global network architecture, which is capable of learning multi-scale features for the graph-based representations. We evaluate the proposed method on challenging benchmark datasets for both 3D hand pose estimation and 3D body pose estimation. Experimental results show that our method achieves state-of-the-art performance on both tasks.

In this paper, we propose a novel top-down instance segmentation framework based on explicit shape encoding, named ESE-Seg. It largely reduces the computational consumption of the instance segmentation by explicitly decoding the multiple object shapes with tensor operations, thus performs the instance segmentation at almost the same speed as the object detection. ESE-Seg is based on a novel shape signature Inner-center Radius (IR), Chebyshev polynomial fitting and the strong modern object detectors. ESE-Seg with YOLOv3 outperforms the Mask R-CNN on Pascal VOC 2012 at mAP^r@0.5 while 7 times faster.

3D object detection and pose estimation from a single image are two inherently ambiguous problems. Oftentimes, objects appear similar from different viewpoints due to shape symmetries, occlusion and repetitive textures. This ambiguity in both detection and pose estimation means that an object instance can be perfectly described by several different poses and even classes. In this work we propose to explicitly deal with these ambiguities. For each object instance we predict multiple 6D pose outcomes to estimate the specific pose distribution generated by symmetries and repetitive textures. The distribution collapses to a single outcome when the visual appearance uniquely identifies just one valid pose. We show the benefits of our approach which provides not only a better explanation for pose ambiguity, but also a higher accuracy in terms of pose estimation.

Fitting model parameters to a set of noisy data points is a common problem in computer vision. In this work, we fit the 6D camera pose to a set of noisy correspondences between the 2D input image and a known 3D environment. We estimate these correspondences from the image using a neural network. Since the correspondences often contain outliers, we utilize a robust estimator such as Random Sample Consensus (RANSAC) or Differentiable RANSAC (DSAC) to fit the pose parameters. When the problem domain, e.g. the space of all 2D-3D correspondences, is large or ambiguous, a single network does not cover the domain well. Mixture of Experts (MoE) is a popular strategy to divide a problem domain among an ensemble of specialized networks, so called experts, where a gating network decides which expert is responsible for a given input. In this work, we introduce Expert Sample Consensus (ESAC), which integrates DSAC in a MoE. Our main technical contribution is an efficient method to train ESAC jointly and end-to-end. We demonstrate experimentally that ESAC handles two real-world problems better than competing methods, i.e. scalability and ambiguity. We apply ESAC to fitting simple geometric models to synthetic images, and to camera re-localization for difficult, real datasets.

We present a new method for finding video CNN architectures that more optimally capture rich spatio-temporal information in videos. Previous work, taking advantage of 3D convolutions, obtained promising results by manually designing CNN video architectures. We here develop a novel evolutionary algorithm that automatically explores models with different types and combinations of layers to jointly learn interactions between spatial and temporal aspects of video representations. We demonstrate the generality of this algorithm by applying it to two meta-architectures. Further, we propose a new component, the iTGM layer, which more efficiently utilizes its parameters to allow learning of space-time interactions over longer time horizons. The iTGM layer is often preferred by the evolutionary algorithm and allows building cost-efficient networks. The proposed approach discovers new diverse and interesting video architectures that were unknown previously. More importantly they are both more accurate and faster than prior models, and outperform the state-of-the-art results on four datasets: Kinetics, Charades, Moments in Time and HMDB. We will open source the code and models, to encourage future model development.

This paper presents a simple method for "do as I do" motion transfer: given a source video of a person dancing, we can transfer that performance to a novel (amateur) target after only a few minutes of the target subject performing standard moves. We approach this problem as video-to-video translation using pose as an intermediate representation. To transfer the motion, we extract poses from the source subject and apply the learned pose-to-appearance mapping to generate the target subject. We predict two consecutive frames for temporally coherent video results and introduce a separate pipeline for realistic face synthesis. Although our method is quite simple, it produces surprisingly compelling results (see video). This motivates us to also provide a forensics tool for reliable synthetic content detection, which is able to distinguish videos synthesized by our system from real data. In addition, we release a first-of-its-kind open-source dataset of videos that can be legally used for training and motion transfer.

In contrast to traditional cameras, whose pixels have a common exposure time, event-based cameras are novel bio-inspired sensors whose pixels work independently and asynchronously output intensity changes (called "events"), with microsecond resolution. Since events are caused by the apparent motion of objects, event-based cameras sample visual information based on the scene dynamics and are, therefore, a more natural fit than traditional cameras to acquire motion, especially at high speeds, where traditional cameras suffer from motion blur. However, distinguishing between events caused by different moving objects and by the camera's ego-motion is a challenging task. We present the first per-event segmentation method for splitting a scene into independently moving objects. Our method jointly estimates the event-object associations (i.e., segmentation) and the motion parameters of the objects (or the background) by maximization of an objective function, which builds upon recent results on event-based motion-compensation. We provide a thorough evaluation of our method on a public dataset, outperforming the state-of-the-art by as much as 10%. We also show the first quantitative evaluation of a segmentation algorithm for event cameras, yielding around 90% accuracy at 4 pixels relative displacement.

Single-image super-resolution aims to generate a high-resolution version of a low-resolution image, which serves as an essential component in many image processing applications. This paper investigates the robustness of deep learning-based super-resolution methods against adversarial attacks, which can significantly deteriorate the super-resolved images without noticeable distortion in the attacked low-resolution images. It is demonstrated that state-of-the-art deep super-resolution methods are highly vulnerable to adversarial attacks. Different levels of robustness of different methods are analyzed theoretically and experimentally. We also present analysis on transferability of attacks, and feasibility of targeted attacks and universal attacks.

Batch normalization (BN) has been very effective for deep learning and is widely used. However, when training with small minibatches, models using BN exhibit a significant degradation in performance. In this paper we study this peculiar behavior of BN to gain a better understanding of the problem, and identify a cause. We propose `EvalNorm' to address the issue by estimating corrected normalization statistics to use for BN during evaluation. EvalNorm supports online estimation of the corrected statistics while the model is being trained, and does not affect the training scheme of the model. As a result, EvalNorm can also be used with existing pre-trained models allowing them to benefit from our method. EvalNorm yields large gains for models trained with smaller batches. Our experiments show that EvalNorm performs 6.18% (absolute) better than vanilla BN for a batchsize of 2 on ImageNet validation set and from 1.5 to 7.0 points (absolute) gain on the COCO object detection benchmark across a variety of setups.

We present a general method to estimate the fundamental matrix from a pair of images under perspective projection without the need for image point correspondences. Our method is particularly well-suited for transmission imaging, where state-of-the-art feature detection and matching approaches generally do not perform well. Estimation of the fundamental matrix plays a central role in auto-calibration methods for reflection imaging. Such methods are currently not applicable to transmission imaging. Furthermore, our method extends an existing technique proposed for reflection imaging which potentially avoids the outlier-prone feature matching step from an orthographic projection model to a perspective model. Our method exploits the idea that under a linear attenuation model line integrals along corresponding epipolar lines are equal if we compute their derivatives in orthogonal direction to their common epipolar plane. We use the fundamental matrix to parametrize this equality. Our method estimates the matrix by formulating a non-convex optimization problem, minimizing an error in our measurement of this equality. We believe this technique will enable the application of the large body of work on image-based camera pose estimation to transmission imaging leading to more accurate and more general motion compensation and auto-calibration algorithms, particularly in medical X-ray and Computed Tomography imaging.

We introduce PlatonicGAN to discover the 3D structure of an object class from an unstructured collection of 2D images, i.e., where no relation between photos is known, except that they are showing instances of the same category. The key idea is to train a deep neural network to generate 3D shapes which, when rendered to images, are indistinguishable from ground truth images (for a discriminator) under various camera poses. Discriminating 2D images instead of 3D shapes allows tapping into unstructured 2D photo collections instead of relying on curated (e.g., aligned, annotated, etc.) 3D data sets. To establish constraints between 2D image observation and their 3D interpretation, we suggest a family of rendering layers that are effectively differentiable. This family includes visual hull, absorption-only (akin to x-ray), and emission-absorption. We can successfully reconstruct 3D shapes from unstructured 2D images and extensively evaluate PlatonicGAN on a range of synthetic and real data sets achieving consistent improvements over baseline methods. We further show that PlatonicGAN can be combined with 3D supervision to improve on and in some cases even surpass the quality of 3D-supervised methods.

In image captioning, the typical attention mechanisms are arduous to identify the equivalent visual signals especially when predicting highly abstract words. This phenomenon is known as the semantic gap between vision and language. This problem can be overcome by providing semantic attributes that are homologous to language. Thanks to the inherent recurrent nature and gated operating mechanism, Recurrent Neural Network (RNN) and its variants are the dominating architectures in image captioning. However, when designing elaborate attention mechanisms to integrate visual inputs and semantic attributes, RNN-like variants become unflexible due to their complexities. In this paper, we investigate a Transformer-based sequence modeling framework, built only with attention layers and feedforward layers. To bridge the semantic gap, we introduce EnTangled Attention (ETA) that enables the Transformer to exploit semantic and visual information simultaneously. Furthermore, Gated Bilateral Controller (GBC) is proposed to guide the interactions between the multimodal information. We name our model as ETA-Transformer. Remarkably, ETA-Transformer achieves state-of-the-art performance on the MSCOCO image captioning dataset. The ablation studies validate the improvements of our proposed modules.

We propose a single-stage detection framework that jointly tackles the problem of multi-scale object detection and class imbalance. Rather than designing deeper networks, we introduce a simple yet effective feature enrichment scheme to produce multi-scale contextual features. We further introduce a cascaded refinement scheme which first instills multi-scale contextual features into the prediction layers of the single-stage detector in order to enrich their discriminative power for multi-scale detection. Second, the cascaded refinement scheme counters the class imbalance problem by refining the anchors and enriched features to improve classification and regression. Experiments are performed on two benchmarks: PASCAL VOC and MS COCO. For a 320x320 input on the MS COCO test-dev, our detector achieves state-of-the-art single-stage detection accuracy with a COCO AP of 33.2 in the case of single-scale inference, while operating at 21 milliseconds on a Titan XP GPU. For a 512x512 input on the MS COCO test-dev, our approach obtains an absolute gain of 1.6% in terms of COCO AP, compared to the best reported single-stage results[5]. Source code and models are available at: https://github.com/Ranchentx/EFGRNet.

Enhancing low light videos, which consists of denoising and brightness adjustment, is an intriguing but knotty problem. Under low light condition, due to high sensitivity camera setting, commonly negligible noises become obvious and severely deteriorate the captured videos. To recover high quality videos, a mass of image/video denoising/enhancing algorithms are proposed, most of which follow a set of simple assumptions about the statistic characters of camera noise, e.g., independent and identically distributed(i.i.d.), white, additive, Gaussian, Poisson or mixture noises. However, the practical noise under high sensitivity setting in real captured videos is complex and inaccurate to model with these assumptions. In this paper, we explore the physical origins of the practical high sensitivity noise in digital cameras, model them mathematically, and propose to enhance the low light videos based on the noise model by using an LSTM-based neural network. Specifically, we generate the training data with the proposed noise model and train the network with the dark noisy video as input and clear-bright video as output. Extensive comparisons on both synthetic and real captured low light videos with the state-of-the-art methods are conducted to demonstrate the effectiveness of the proposed method.

Neural networks are vulnerable to adversarial examples, malicious inputs crafted to fool trained models. Adversarial examples often exhibit black-box transfer, meaning that adversarial examples for one model can fool another model. However, adversarial examples are typically overfit to exploit the particular architecture and feature representation of a source model, resulting in sub-optimal black-box transfer attacks to other target models. We introduce the Intermediate Level Attack (ILA), which attempts to fine-tune an existing adversarial example for greater black-box transferability by increasing its perturbation on a pre-specified layer of the source model, improving upon state-of-the-art methods. We show that we can select a layer of the source model to perturb without any knowledge of the target models while achieving high transferability. Additionally, we provide some explanatory insights regarding our method and the effect of optimizing for adversarial examples using intermediate feature maps.

Multi-view shape descriptors obtained from various 2D images are commonly adopted in 3D shape retrieval. One major challenge is that significant shape information are discarded during 2D view rendering through projection. In this paper, we propose a convolutional neural network based method, CenterNet, to enhance each individual 2D view using its neighboring ones. By exploiting cross-view correlations, CenterNet learns how adjacent views can be maximally incorporated for an enhanced 2D representation to effectively describe shapes. We observe that a very small amount of, e.g., six, enhanced 2D views, are already sufficient for a panoramic shape description. Thus, by simply aggregating features from six enhanced 2D views, we arrive at a highly compact yet discriminative shape descriptor. The proposed shape descriptor significantly outperforms state-of-the-art 3D shape retrieval methods on the ModelNet and ShapeNetCore55 benchmarks, and also exhibits robustness against object occlusion.

Monocular depth prediction plays a crucial role in understanding 3D scene geometry. Although recent methods have achieved impressive progress in evaluation metrics such as the pixel-wise relative error, most methods neglect the geometric constraints in the 3D space. In this work, we show the importance of the high-order 3D geometric constraints for depth prediction. By designing a loss term that enforces one simple type of geometric constraints, namely, virtual normal directions determined by randomly sampled three points in the reconstructed 3D space, we can considerably improve the depth prediction accuracy. Furthermore, we can not only predict accurate depth but also achieve high-quality other 3D information from the depth without retraining new parameters, Significantly, the byproduct of this predicted depth being sufficiently accurate is that we are now able to recover good 3D structures of the scene such as the point cloud and surface normal directly from the depth, eliminating the necessity of training new sub-models as was previously done. Experiments on two challenging benchmarks: NYU Depth-V2 and KITTI demonstrate the effectiveness of our method and state-of-the-art performance.

We present a conceptually simple yet effective algorithm to detect wireframes in a given image. Compared to the previous methods which first predict an intermediate heat map and then extract straight lines with heuristic algorithms, our method is end-to-end trainable and can directly output a vectorized wireframe that contains semantically meaningful and geometrically salient junctions and lines. To better understand the quality of the outputs, we propose a new metric for wireframe evaluation that penalizes overlapped line segments and incorrect line connectivities. We conduct extensive experiments and show that our method significantly outperforms the previous state-of-the-art wireframe and line extraction algorithms. We hope our simple approach can be served as a baseline for future wireframe parsing studies. Code has been made publicly available at https://github.com/zhou13/lcnn.

Event cameras are vision sensors that record asynchronous streams of per-pixel brightness changes, referred to as "events". They have appealing advantages over frame based cameras for computer vision, including high temporal resolution, high dynamic range, and no motion blur. Due to the sparse, non-uniform spatio-temporal layout of the event signal, pattern recognition algorithms typically aggregate events into a grid-based representation and subsequently process it by a standard vision pipeline, e.g., Convolutional Neural Network (CNN). In this work, we introduce a general framework to convert event streams into grid-based representations by means of strictly differentiable operations. Our framework comes with two main advantages: (i) allows learning the input event representation together with the task dedicated network in an end to end manner, and (ii) lays out a taxonomy that unifies the majority of extant event representations in the literature and identifies novel ones. Empirically, we show that our approach to learning the event representation end-to-end yields an improvement of approximately 12% on optical flow estimation and object recognition over state-of-the-art methods.