ECCV 2020 Papers

Where people look when watching videos is believed to be heavily influenced by temporal patterns. In this work, we test this assumption by quantifying to which extent gaze on recent video saliency benchmarks can be predicted by a static baseline model. On the recent LEDOV dataset, we find that at least 75% of the explainable information as defined by a gold standard model can be explained using static features. Our baseline model ``DeepGaze MR'' even outperforms state-of-the-art video saliency models, despite deliberately ignoring all temporal patterns. Visual inspection of our static baseline’s failure cases shows that clear temporal effects on human gaze placement exist, but are both rare in the dataset and not captured by any of the recent video saliency models. To focus the development of video saliency models on better capturing temporal effects we construct a meta-dataset consisting of those examples requiring temporal information."

Articulations, Shapes and Objects for 3D Hand Pose Estimation under Hand-Object Interaction","We study how well different types of approaches generalise in the task of 3D hand pose estimation under single hand scenarios and hand-object interaction. We show that the accuracy of state-of-the-art methods can drop, and that they fail mostly on poses absent from the training set. Unfortunately, since the space of hand poses is highly dimensional, it is inherently not feasible to cover the whole space densely, despite recent efforts in collecting large-scale training datasets. This sampling problem is even more severe when hands are interacting with objects and/or inputs are RGB rather than depth images, as RGB images also vary with lighting conditions and colors. To address these issues, we designed a public challenge (HANDS'19) to evaluate the abilities of current 3D hand pose estimators~(HPEs) to interpolate and extrapolate the poses of a training set. More exactly, HANDS'19 is designed (a) to evaluate the influence of both depth and color modalities on 3D hand pose estimation, under the presence or absence of objects; (b) to assess the generalisation abilities w.r.t. four main axes: shapes, articulations, viewpoints, and objects; (c) to explore the use of a synthetic hand models to fill the gaps of current datasets. Through the challenge, the overall accuracy has dramatically improved over the baseline, especially on extrapolation tasks, from 27mm to 13mm mean joint error. Our analyses highlight the impacts of: ensemble approaches, the use of a parametric 3D hand model (MANO), and different HPE methods/backbones."

We introduce a method to convert stereo 360 (omnidirectional stereo) imagery into a layered, multi-sphere image representation for six degree-of-freedom (6DoF) rendering. Stereo 360 imagery can be captured from multi-camera systems for virtual reality (VR) rendering, but lacks motion parallax and correct-in-all-directions disparity cues. Together, these can quickly lead to VR sickness when viewing content. One solution is to try and generate a format suitable for 6DoF rendering, such as by estimating depth. However, this raises questions as to how to handle disoccluded regions in dynamic scenes. Our approach is to simultaneously learn depth and blending weights via a multi-sphere image representation, which can be rendered with correct 6DoF disparity and motion parallax in VR. This significantly improves comfort for the viewer, and can be inferred and rendered in real time on modern GPU hardware. Together, these move towards making VR video a more comfortable immersive medium."

Feature distillation is an effective way to improve the performance for a smaller student model, which has fewer parameters and lower computation cost compared to the larger teacher model. Unfortunately, there is a common obstacle $-$ the gap in semantic feature structure between the intermediate features of teacher and student. The classic scheme prefers to transform intermediate features by adding the adaptation module, such as naive convolutional, attention-based or more complicated one. However, this introduces two problems: a) The adaptation module brings more parameters into training. b) The adaptation module with random initialization or special transformation isn't friendly for distilling a pre-trained student. In this paper, we present Matching Guided Distillation (MGD) as an efficient and parameter-free manner to solve these problems. The key idea of MGD is to pose matching the teacher channels with students' as an assignment problem. We compare three solutions of the assignment problem to reduce channels from teacher features with partial distillation loss. The overall training takes a coordinate-descent approach between two optimization objects $-$ assignments update and parameters update. Since MGD only contains normalization or pooling operations with negligible computation cost, it is flexible to plug into network with other distillation methods. The project site is http://kaiyuyue.com/mgd ."

Object recognition has seen significant progress in the image domain, with focus primarily on 2D perception. We propose to leverage existing large-scale datasets of 3D models to understand the underlying 3D structure of objects seen in an image by constructing a CAD-based representation of the objects and their poses. We present Mask2CAD, which jointly detects objects in real-world images and for each detected object, optimizes for the most similar CAD model and its pose.

Recent end-to-end trainable methods for scene text spotting, integrating detection and recognition, showed much progress. However, most of the current arbitrary-shape scene text spotters use region proposal networks (RPN) to produce proposals. RPN relies heavily on manually designed anchors and its proposals are represented with axis-aligned rectangles. The former presents difficulties in handling text instances of extreme aspect ratios or irregular shapes, and the latter often includes multiple neighboring instances into a single proposal, in cases of densely oriented text. To tackle these problems, we propose Mask TextSpotter v3, an end-to-end trainable scene text spotter that adopts a Segmentation Proposal Network (SPN) instead of an RPN. Our SPN is anchor-free and gives accurate representations of arbitrary-shape proposals. It is therefore superior to RPN in detecting text instances of extreme aspect ratios or irregular shapes. Furthermore, the accurate proposals produced by SPN allow masked RoI features to be used for decoupling neighboring text instances. As a result, our Mask TextSpotter v3 can handle text instances of extreme aspect ratios or irregular shapes, and its recognition accuracy won't be affected by nearby text or background noise. Specifically, we outperform state-of-the-art methods by 21.9 percent on the Rotated ICDAR 2013 dataset (rotation robustness), 5.9 percent on the Total-Text dataset (shape robustness), and achieve state-of-the-art performance on the MSRA-TD500 dataset (aspect ratio robustness). Code is available at: https://github.com/MhLiao/MaskTextSpotterV3 "

In this paper, as we aim to construct a semi-supervised learning algorithm, we exploit the characteristics of the Deep Convolutional Networks to provide, for an input image, both an embedding descriptor and a prediction. The unlabeled data is combined with the labeled one in order to provide synthetic data, which describes better the input space. The network is asked to provide a large margin between clusters, while new data is self-labeled by the distance to class centroids, in the embedding space. The method is tested on standard benchmarks for semi--supervised learning, where it matches state of the art performance and on the problem of face expression recognition where it increases the accuracy by a noticeable margin."

We propose an unsupervised learning framework with the pretext task of finding dense correspondences between point cloud shapes from the same category based on the cycle-consistency formulation. In order to learn discriminative pointwise features from point cloud data, we incorporate in the formulation a regularization term based on Sinkhorn normalization to enhance the learned pointwise mappings to be as bijective as possible. Besides, a random rigid transform of the source shape is introduced to form a triplet cycle to improve the model's robustness against perturbations. Comprehensive experiments demonstrate that the learned pointwise features through our framework benefits various point cloud analysis tasks, e.g. partial shape registration and keypoint transfer. We also show that the learned pointwise features can be leveraged by supervised methods to improve the part segmentation performance with either the full training dataset or just a small portion of it."

Learning-based methods produce remarkable results on single image depth tasks when trained on well-established benchmarks, however, there is a large gap from these benchmarks to real-world performance that is usually obscured by the common practice of fine-tuning on the target dataset. We introduce a new depth dataset that is an order of magnitude larger than previous offerings, but more importantly, contains an unprecedented gamut of locations, camera models and scene types while offering metric depth (not just up-to-scale). Additionally, we investigate the problem of training single image depth networks using images captured with many different cameras, validating an existing approach and proposing a simpler alternative. With our contributions we achieve excellent results on challenging benchmarks before fine-tuning, and set the state of the art on the popular KITTI dataset after fine-tuning."

Object detection has witnessed significant progress by relying on large, manually annotated datasets. Annotating such datasets is highly time consuming and expensive, which motivates the development of weakly supervised and few-shot object detection methods. However, these methods largely underperform with respect to their strongly supervised counterpart, as weak training signals mph{often} result in partial or oversized detections. Towards solving this problem we introduce, for the first time, an online annotation module (OAM) that learns to generate a many-shot set of mph{reliable} annotations from a larger volume of weakly labelled images. Our OAM can be jointly trained with any fully supervised two-stage object detection method, providing additional training annotations on the fly. This results in a fully end-to-end strategy that only requires a low-shot set of fully annotated images. The integration of the OAM with Fast(er) R-CNN improves their performance by $17\%$ mAP, $9\%$ AP50 on PASCAL VOC 2007 and MS-COCO benchmarks, and significantly outperforms competing methods using mixed supervision."

Although deep convolutional neural network based face recognition system has achieved remarkable success, it is susceptible to adversarial images: carefully constructed imperceptible perturbations can easily mislead deep neural networks. A recent study has shown that in addition to regular off-manifold adversarial images, there are also adversarial images on the manifold. In this paper, we propose adversarial variational autoencoder (A-VAE), a novel framework to tackle both types of attacks. We hypothesize that both off-manifold and on-manifold attacks move the image away from the high probability region of image manifold. We utilize variational autoencoder (VAE) to estimate the lower bound of the log-likelihood of image and explore to project the input images back into the high probability regions of image manifold again. At inference time, our model synthesizes multiple similar realizations of a given image by random sampling, then the nearest neighbor of the given image is selected as the final input of the face recognition model. We also use adversarial training to enhance the robustness of our model against adversarial perturbations. As a preprocessing operation, our method is attack-agnostic and can adapt to a wide range of resolutions. The experimental results on LFW demonstrate that our method achieves state-of-the-art defense success rate against conventional off-manifold attacks such as FGSM, PGD, and C\&W under both grey-box and white-box settings, and even on-manifold attack."

Depth data provide geometric information that can bring progress in RGB-D scene parsing tasks. Several recent works propose RGB-D convolution operators that construct receptive fields along the depth-axis to handle 3D neighborhood relations between pixels. However, these methods pre-define depth receptive fields by hyperparameters, making them rely on parameter selection. In this paper, we propose a novel operator called malleable 2.5D convolution to learn the receptive field along the depth-axis. A malleable 2.5D convolution has one or more 2D convolution kernels. Our method assigns each pixel to one of the kernels or none of them according to their relative depth differences, and the assigning process is formulated as a differentiable form so that it can be learnt by gradient descent. The proposed operator runs on standard 2D feature maps and can be seamlessly incorporated into pre-trained CNNs. We conduct extensive experiments on two challenging RGB-D semantic segmentation dataset NYUDv2 and Cityscapes to validate the effectiveness and the generalization ability of our method."

We present a systematic study of adversarial attacks on state-of-the-art object detection frameworks. Using standard detection datasets, we train patterns that suppress the objectness scores produced by a range of commonly used detectors, and ensembles of detectors. Through extensive experiments, we benchmark the effectiveness of adversarially trained patches under both white-box and black-box settings, and quantify transferability of attacks between datasets, object classes, and detector models. Finally, we present a detailed study of physical world attacks using printed posters and wearable clothes, and rigorously quantify the performance of such attacks with different metrics."

To tackle increasingly complex tasks, it has become an essential ability of neural networks to learn abstract representations. These task-specific representations and, particularly, the invariances they capture turn neural networks into black box models that lack interpretability. To open such a black box, it is, therefore, crucial to uncover the different semantic concepts a model has learned as well as those that it has learned to be invariant to. We present an approach based on INNs that (i) recovers the task-specific, learned invariances by disentangling the remaining factor of variation in the data and that (ii) invertibly transforms these recovered invariances combined with the model representation into an equally expressive one with accessible semantic concepts. As a consequence, neural network representations become understandable by providing the means to (i) expose their semantic meaning, (ii) semantically modify a representation, and (iii) visualize individual learned semantic concepts and invariances. Our invertible approach significantly extends the abilities to understand black box models by enabling post-hoc interpretations of state-of-the-art networks without compromising their performance."

Local features such as SIFT and its affine and learned variants provide region-to-region rather than point-to-point correspondences. It has recently been exploited to create new minimal solvers for classical problems such as homography, essential and fundamental matrix estimation. The main argument in favor of such solvers is that their minimal sample size is smaller, e.g., only two instead of four matches are required to estimate a homography. Works proposing such solvers often claim a significant improvement in run-time thanks to fewer RANSAC iterations. We show that this argument is not valid in practice if the solvers are used naively as noise in the local feature geometries often causes RANSAC to fail to find an accurate solution. To overcome this, we propose guidelines for effective use of region-to-region matches in the course of a full model estimation pipeline. We propose a method for refining the local feature geometries by symmetric intensity-based matching, combine uncertainty propagation inside RANSAC with preemptive model verification, show a general scheme for computing uncertainty of minimal solvers results, and adapt the sample cheirality check for homography estimation to region-to-region correspondences. Our experiments show that affine solvers can achieve accuracy comparable to point-based solvers at faster run-times when following our guidelines. "

In this paper, we argue about the importance of considering task interactions at multiple scales when distilling task information in a multi-task learning setup. In contrast to common belief, we show that tasks with high affinity at a certain scale are not guaranteed to retain this behaviour at other scales, and vice versa. We propose a novel architecture, namely MTI-Net, that builds upon this finding in three ways. First, it explicitly models task interactions at every scale via a multi-scale multi-modal distillation unit. Second, it propagates distilled task information from lower to higher scales via a feature propagation module. Third, it aggregates the refined task features from all scales via a feature aggregation unit to produce the final per-task predictions.

Clustering is a fundamental task in unsupervised learning that depends heavily on the data representation that is used. Deep generative models have appeared as a promising tool to learn informative low-dimensional data representations. We propose Matching Priors and Conditionals for Clustering (MPCC), a GAN-based model with an encoder to infer latent variables and cluster categories from data, and a flexible decoder to generate samples from a conditional latent space. With MPCC we demonstrate that a deep generative model can be competitive/superior against discriminative methods in clustering tasks surpassing the state of the art over a diverse set of benchmark datasets. Our experiments show that adding a learnable prior and augmenting the number of encoder updates improve the quality of the generated samples, obtaining an inception score of $9.49 \pm 0.15$ and improving the Fr\'echet inception distance over the state of the art by a $46.9\%$ in CIFAR10."

We address the weakly supervised video highlight detection problem for learning to detect segments that are more attractive in training videos given their video event label but without expensive supervision of manually annotating highlight segments. While manually averting localizing highlight segments, weakly supervised modeling is challenging, as a video in our daily life could contain highlight segments with multiple event types, e.g., skiing and surfing. In this work, we propose casting weakly supervised video highlight detection modeling for a given specific event as a multiple instance ranking network (MINI-Net) learning. We consider each video as a bag of segments, and therefore, the proposed MINI-Net learns to enforce a higher highlight score for a positive bag that contains highlight segments of a specific event than those for negative bags that are irrelevant. In particular, we form a max-max ranking loss to acquire a reliable relative comparison between the most likely positive segment instance and the hardest negative segment instance. With this max-max ranking loss, our MINI-Net effectively leverages all segment information to acquire a more distinct video feature representation for localizing the highlight segments of a specific event in a video. The extensive experimental results on three challenging public benchmarks clearly validate the efficacy of our multiple instance ranking approach for solving the problem."

The synthesis of natural emotional reactions is an essentialcriteria in vivid talking-face video generation. This criteria is nevertheless seldom taken into consideration in previous works due to the absence of a large-scale, high-quality emotional audio-visual dataset. To address this issue, we build the Multi-view Emotional Audio-visual Dataset(MEAD) which is a talking-face video corpus featuring 60 actors and actresses talking with 8 different emotions at 3 different intensity levels. High-quality audio-visual clips are captured at 7 different view angles in a strictly-controlled environment. Together with the dataset, we release an emotional talking-face generation baseline which enables the manipulation of both emotion and its intensity. Our dataset will be made public and could benefit a number of different research fields including conditional generation, cross-modal understanding and expression recognition."