ECCV 2020 Papers

The abundance of multimodal data (e.g. social media posts) has inspired interest in cross-modal retrieval methods. Popular approaches rely on a variety of metric learning losses, which prescribe what the proximity of image and text should be, in the learned space. However, most prior methods have focused on the case where image and text convey redundant information; in contrast, real-world image-text pairs convey complementary information with little overlap. Further, images in news articles and media portray topics in a visually diverse fashion; thus, we need to take special care to ensure a meaningful image representation. We propose novel within-modality losses which encourage semantic coherency in both the text and image subspaces, which does not necessarily align with visual coherency. Our method ensures that not only are paired images and texts close, but the expected image-image and text-text relationships are also observed. Our approach improves the results of cross-modal retrieval on four datasets compared to five baselines."

Low-resolution person re-identification (LR re-id) is a challenging task with low-resolution probes and high-resolution gallery images. To address the resolution mismatch, existing methods typically recover missing details for low-resolution probes by super-resolution. However, they usually pre-specify fixed scale factors for all images, and ignore the fact that choosing a preferable scale factor for certain image content probably greatly benefits the identification. In this paper, we propose a novel Prediction, Recovery and Identification (PRI) model for LR re-id, which adaptively recovers missing details by predicting a preferable scale factor based on the image content. To deal with the lack of ground-truth optimal scale factors, our model contains a self-supervised scale factor metric that automatically generates dynamic soft labels. The generated labels indicate probabilities that each scale factor is optimal, which are used as guidance to enhance the content-aware scale factor prediction. Consequently, our model can more accurately predict and recover the content-aware details, and achieve state-of-the-art performances on four LR re-id datasets."

To what extent are two images picturing the same 3D surfaces? Even when this is a known scene, the answer typically requires an expensive search across scale space, with matching and geometric verification of large sets of local features. This expense is further multiplied when a query image is evaluated against a gallery, e.g. in visual relocalization. While we don't obviate the need for geometric verification, we propose an interpretable image-embedding that cuts the search in scale space to essentially a lookup.

Data poisoning attacks on machine learning models have attracted much recent attention, wherein poisoning samples are injected at the training phase to achieve adversarial goals at test time. Although existing poisoning techniques prove to be effective in various scenarios, they rely on certain assumptions on the adversary knowledge and capability to ensure efficacy, which may be unrealistic in practice. This paper presents a new, practical targeted poisoning attack method on neural networks in vision domain, namely BlackCard. BlackCard possesses a set of critical properties for ensuring attacking efficacy in practice, which has never been simultaneously achieved by any existing work, including knowledge-oblivious, clean-label, and clean-test. Importantly, we show that the effectiveness of BlackCard can be intuitively guaranteed by a set of analytical reasoning and observations, through exploiting an essential characteristic of gradient-descent optimization which is pervasively adopted in DNN models. We evaluate the efficacy of BlackCard for generating targeted poisoning attacks via extensive experiments using various datasets and DNN models. Results show that BlackCard is effective with a rather high success rate while preserving all the claimed properties."

When the training data are maliciously tampered, the predictions of the acquired deep neural network (DNN) can be manipulated by an adversary known as the Trojan attack (or poisoning backdoor attack). The lack of robustness of DNNs against Trojan attacks could significantly harm real-life machine learning (ML) systems in downstream applications, therefore posing widespread concern to their trustworthiness. In this paper, we study the problem of the Trojan network (TrojanNet) detection in the data-scarce regime, where only the weights of a trained DNN are accessed by the detector. We first propose a data-limited TrojanNet detector (TND), when only a few data samples are available for TrojanNet detection. We show that an effective data-limited TND can be established by exploring connections between Trojan attack and prediction-evasion adversarial attacks including per-sample attack as well as all-sample universal attack. In addition, we propose a data-free TND, which can detect a TrojanNet without accessing any data samples. We show that such a TND can be built by leveraging the internal response of hidden neurons, which exhibits the Trojan behavior even at random noise inputs. The effectiveness of our proposals is evaluated by extensive experiments under different model architectures and datasets including CIFAR-10, GTSRB, and ImageNet."

Deep learning-based image denoising approaches have been extensively studied in recent years, prevailing in many public benchmark datasets. However, the stat-of-the-art networks are computationally too expensive to be directly applied on mobile devices. In this work, we propose a light-weight, efficient neural network-based raw image denoiser that runs smoothly on mainstream mobile devices, and produces high quality denoising results. Our key insights are twofold:(1) by measuring and estimating sensor noise level, a smaller network trained on synthetic sensor-specific data can out-perform larger ones trained on general data;(2) the large noise level variation under different ISO settings can be removed by a novel k-Sigma Transform, allowing a small network to efficiently handle a wide range of noise levels. We conduct extensive experiments to demonstrate the efficiency and accuracy of our approach. Our proposed mobile-friendly denoising model runs at ~70 milliseconds per megapixel on Qualcomm Snapdragon 855 chipset, and it is the basis of the night shot feature of several flagship smartphones released in 2019. "

One-shot NAS method has attracted much interest from the research community due to its remarkable training efficiency and capacity to discover high performance models. However, the search spaces of previous one-shot based works usually relied on hand-craft design and were short for flexibility on the network topology. In this work, we try to enhance the one-shot NAS by exploring high-performing network architectures in our large-scale Topology Augmented Search Space (i.e., over 3.4×10^10 different topological structures). Specifically, the difficulties for architecture searching in such a complex space has been eliminated by the proposed stabilized share-parameter proxy, which employs Stochastic Gradient Langevin Dynamics to enable fast shared parameter sampling, so as to achieve stabilized measurement of architecture performance even in search space with complex topological structures. The proposed method, namely Stablized Topological Neural Architecture Search (ST-NAS), achieves state-of-the-art performance under Multiply-Adds (MAdds) constraint on ImageNet. Our lite model ST-NAS-A achieves 76.4% top-1 accuracy with only 326M MAdds. Our moderate model ST-NAS-B achieves 77.9% top-1 accuracy just required 503M MAdds. Both of our models offer superior performances in comparison to other concurrent works on one-shot NAS."

Quantization plays an important role in the energy-efficient deployment of Deep Neural Networks (DNNs) on resource-limited devices. Post-training quantization is highly desirable since it does not require retraining or access to the full training dataset. The well-established uniform scheme for post-training quantization achieves satisfactory results by converting DNNs from full-precision to 8-bit fixed-point integers. However, it suffers from significant performance degradation when quantizing to lower bit-widths. In this paper, we propose a PieceWise Linear Quantization (PWLQ) scheme to enable accurate approximation for tensor values that have bell-shaped distributions with long tails. Our approach breaks the entire quantization range into non-overlapping regions for each tensor, with each region being assigned an equal number of quantization levels. Optimal breakpoints that divide the entire range are found by minimizing the quantization error. Compared to state-of-the-art post-training quantization methods, experimental results show that our proposed method achieves superior performance on image classification, semantic segmentation, and object detection with minor overhead. "

Most of the recent deep learning-based 3D human pose and mesh estimation methods regress the pose and shape parameters of human mesh models, such as SMPL and MANO, from an input image. The first weakness of these methods is the overfitting to image appearance, due to the domain gap between the training data captured from controlled settings such as a lab, and in-the-wild data in inference time. The second weakness is that the estimation of the pose parameters is quite challenging due to the representation issues of 3D rotations. To overcome the above weaknesses, we propose Pose2Mesh, a novel graph convolutional neural network (GraphCNN)-based system that estimates the 3D coordinates of human {m mesh vertices} directly from the {m 2D human pose}. The 2D human pose as an input provides essential human body articulation information without image appearance. Also, the proposed system avoids the representation issues, while fully exploiting the mesh topology using GraphCNN in a coarse-to-fine manner. We show that our Pose2Mesh significantly outperforms the previous 3D human pose and mesh estimation methods on various benchmark datasets. The codes are publicly available ootnote{\url{https://github.com/hongsukchoi/Pose2Mesh_RELEASE}}."

Object pose increases interclass object variance which makes object recognition from 2D images harder. To render a classifier robust to pose variations, most deep neural networks try to eliminate the influence of pose by using large datasets with many poses for each class. Here, we propose a different approach: a class-agnostic object pose transformation network (OPT-Net) can transform an image along 3D yaw and pitch axes to synthesize additional poses continuously. Synthesized images lead to better training of an object classifier. We design a novel eliminate-add structure to explicitly disentangle pose from object identity: first ‘eliminate’ pose information of the input image and then ‘add’ target pose information (regularized as continuous variables) to synthesize any target pose. We trained OPT-Net on images of toy vehicles shot on a turntable from the iLab-20M dataset. After training on unbalanced discrete poses (5 classes with 6 poses per object instance, plus 5 classes with only 2 poses), we show that OPT-Net can synthesize balanced continuous new poses along yaw and pitch axes with high quality. Training a ResNet-18 classifier with original plus synthesized poses improves mAP accuracy by 9% over training on original poses only. Further, the pre-trained OPT-Net can generalize to new object classes, which we demonstrate on both iLab-20M and RGB-D. We also show that the learned features can generalize to ImageNet. (The code is released at this URL)"

Human-Object Interaction (HOI) detection is important in human-centric scene understanding. Existing works typically assume that the same verb in different HOI categories has similar visual characteristics, while ignoring the diverse semantic meanings of the verb. To address this issue, in this paper, we propose a novel Polysemy Deciphering Network (PD-Net), which decodes the visual polysemy of verbs for HOI detection in three ways. First, PD-Net augments human pose and spatial features for HOI detection using language priors, enabling the verb classifiers to receive language hints that reduce the intra-class variation of the same verb. Second, we introduce a novel Polysemy Attention Module (PAM) that guides PD-Net to make decisions based on more important feature types according to the language priors. Finally, the above two strategies are applied to two types of classifiers for verb recognition, i.e., object-shared and object-specific verb classifiers, whose combination further relieves the verb polysemy problem. By deciphering the visual polysemy of verbs, we achieve the best performance on both HICO-DET and V-COCO datasets. In particular, PD-Net outperforms state-of-the-art approaches by 3.81% mAP in the Known-Object evaluation mode of HICO-DET. Code of PD-Net will be released at https://github.com/MuchHair/PD-Net."

Lane detection is a fundamental yet challenging task in autonomous driving and intelligent traffic systems due to perspective projection and occlusion. Most of previous methods utilize semantic segmentation to identify the regions of traffic lanes in an image, and then adopt some curve-fitting method to reconstruct the lanes. In this work, we propose to use polynomial curves to represent traffic lanes and then propose a novel polynomial regression network (PRNet) to directly predict them, where semantic segmentation is not involved. Specifically, PRNet consists of one major branch and two auxiliary branches: (1) polynomial regression to estimate the polynomial coefficients of lanes, (2) initialization classification to detect the initial retrieval point of each lane, and (3) height regression to determine the ending point of each lane. Through the cooperation of three branches, PRNet can detect variable-number of lanes and is highly effective and efficient. We experimentally evaluate the proposed PRNet on two popular benchmark datasets: TuSimple and CULane. The results show that our method significantly outperforms the previous state-of-the-art methods in terms of both accuracy and speed. "

We merge by generalization two principles of passive optical sensing of motion. One is common spatially resolved imaging, where motion induces temporal readout changes at high-contrast spatial features, as used in traditional optical-flow. The other is the polarization compass, where axial rotation induces temporal readout changes due to the change of incoming polarization angle, relative to the camera frame. The latter has traditionally been modeled for uniform objects. This merger generalizes the brightness constancy assumption and optical-flow, to handle polarization. It also generalizes the polarization compass concept to handle arbitrarily textured objects. This way, scene regions having partial polarization contribute to motion estimation, irrespective of their texture and non-uniformity. As an application, we derive and demonstrate passive sensing of differential ego-rotation around the camera optical axis. "

A polarization camera has great potential for 3D reconstruction since the angle of polarization (AoP) of reflected light is related to an object's surface normal. In this paper, we propose a novel 3D reconstruction method called Polarimetric Multi-View Inverse Rendering (Polarimetric MVIR) that effectively exploits geometric, photometric, and polarimetric cues extracted from input multi-view color polarization images. We first estimate camera poses and an initial 3D model by geometric reconstruction with a standard structure-from-motion and multi-view stereo pipeline. We then refine the initial model by optimizing photometric and polarimetric rendering errors using multi-view RGB and AoP images, where we propose a novel polarimetric rendering cost function that enables us to effectively constrain each estimated surface vertex's normal while considering four possible ambiguous azimuth angles revealed from the AoP measurement. Experimental results using both synthetic and real data demonstrate that our Polarimetric MVIR can reconstruct a detailed 3D shape without assuming a specific polarized reflection depending on the material."

A key step in any scanning-based asset creation workflow is to convert unordered point clouds to a surface. Classical methods (e.g. Poisson reconstruction) start to degrade in the presence of noisy and partial scans. Hence, deep learning based methods have recently been proposed to produce complete surfaces, even from partial scans. However, such data-driven methods struggle to generalize to new shapes with large geometric and topological variations. We present Points2Surf, a novel patch-based learning framework that produces accurate surfaces directly from raw scans without normals. Learning a prior over a combination of detailed local patches and coarse global information improves generalization performance and reconstruction accuracy. Our extensive comparison on both synthetic and real data demonstrates a clear advantage of our method over state-of-the-art alternatives on previously unseen classes (on average, Points2Surf brings down reconstruction error by 30% over SPR and by 270%+ over deep learning based SotA methods) at the cost of longer computation times and a slight increase in small-scale topological noise in some cases. Our source code, pre-trained model, and dataset are available at: https://github.com/ErlerPhilipp/points2surf"

This work considers a new task in geometric deep learning: generating a triangulation among a set of points in 3D space. We present PointTriNet, a differentiable and scalable approach enabling point set triangulation as a layer in 3D learning pipelines. The method iteratively applies two neural networks: a classification network predicts whether a candidate triangle should appear in the triangulation, while a proposal network suggests additional candidates. Both networks are structured as PointNets over nearby points and triangles, using a novel triangle-relative input encoding. Since these learning problems operate on local geometric data, our method is efficient and scalable, and generalizes to unseen shape categories. Our networks are trained in an unsupervised manner from a collection of shapes represented as point clouds. We demonstrate the effectiveness of this approach for classical meshing tasks, robustness to outliers, and as a component in end-to-end learning systems."

We propose a novel end-to-end deep scene flow model, called PointPWC-Net, that directly processes 3D point cloud scenes with large motions in a coarse-to-fine fashion. Flow computed at the coarse level is upsampled and warped to a finer level, enabling the algorithm to accommodate for large motion without a prohibitive search space. We introduce novel cost volume, upsampling, and warping layers to efficiently handle 3D point cloud data. Unlike traditional cost volumes that require exhaustively computing all the cost values on a high-dimensional grid, our point-based formulation discretizes the cost volume onto input 3D points, and a PointConv operation efficiently computes convolutions on the cost volume. Experiment results on FlyingThings3D and KITTI outperform the state-of-the-art by a large margin. We further explore novel self-supervised losses to train our model and achieve comparable results to state-of-the-art trained with supervised loss. Without any fine-tuning, our method also shows great generalization ability on the KITTI Scene Flow 2015 dataset, outperforming all previous methods."

This paper introduces data augmentation for point clouds by interpolation between examples. Data augmentation by interpolation has shown to be a simple and effective approach in the image domain. Such a mixup is however not directly transferable to point clouds, as we do not have a one-to-one correspondence between the points of two different objects. In this paper, we define data augmentation between point clouds as a shortest path linear interpolation. To that end, we introduce PointMixup, an interpolation method that generates new examples through an optimal assignment of the path function between two point clouds. We prove that our PointMixup finds the shortest path between two point clouds and that the interpolation is assignment invariant and linear. With the definition of interpolation, PointMixup allows to introduce strong interpolation-based regularizers such as mixup and manifold mixup to the point cloud domain. Experimentally, we show the potential of PointMixup for point cloud classification, especially when examples are scarce, as well as increased robustness to noise and geometric transformations to points. The code for PointMixup and the experimental details are publicly available."

Arguably one of the top success stories of deep learning is transfer learning. The finding that pre-training a network on a rich source set (g, ImageNet) can help boost performance once fine-tuned on a usually much smaller target set, has been instrumental to many applications in language and vision. Yet, very little is known about its usefulness in 3D point cloud understanding. We see this as an opportunity considering the effort required for annotating data in 3D. In this work, we aim at facilitating research on 3D representation learning. Different from previous works, we focus on high-level scene understanding tasks. To this end, we select a suit of diverse datasets and tasks to measure the effect of unsupervised pre-training on a large source set of 3D scenes. Our findings are extremely encouraging: using a unified triplet of architecture, source dataset, and contrastive loss for pre-training, we achieve improvement over recent best results in segmentation and detection across 6 different benchmarks for indoor and outdoor, real and synthetic datasets -- demonstrating that the learned representation can generalize across domains. Furthermore, the improvement was similar to supervised pre-training, suggesting that future efforts should favor scaling data collection over more detailed annotation. We hope these findings will encourage more research on unsupervised pretext task design for 3D deep learning."