ECCV 2020 Papers

By adding human-imperceptible noise to clean images, the resultant adversarial examples can fool other unknown models. Features of a pixel extracted by deep neural networks (DNNs) are influenced by its surrounding regions, and different DNNs generally focus on different discriminative regions in recognition. Motivated by this, we propose a \underline{patch}-wise iterative algorithm -- a black-box attack towards mainstream normally trained and defense models, which differs from the existing attack methods manipulating \underline{pixel}-wise noise. In this way, without sacrificing the performance of the substitute model, our adversarial examples can have strong transferability. Specifically, we introduce an amplification factor to the step size in each iteration, and one pixel's overall gradient overflowing the $psilon$-constraint is properly assigned to its surrounding regions by a projection kernel. Our method can be generally integrated to any gradient-based attack methods. Compared with the current state-of-the-art attacks, we significantly improve the success rate by 9.2\% for defense models and 3.7\% for normally trained models on average. Our anonymous code is available at \url{http://tiny.cc/da7vkz}"

Cameras are prevalent in our daily lives, and enable many useful systems built upon computer vision technologies such as smart cameras and home robots for service applications. However, there is also an increasing societal concern as the captured images/videos may contain privacy-sensitive information (e.g., face identity). We propose a novel face identity transformer which enables automated photo-realistic password-based anonymization and deanonymization of human faces appearing in visual data. Our face identity transformer is trained to (1) remove face identity information after anonymization, (2) recover the original face when given the correct password, and (3) return a wrong---but photo-realistic---face given a wrong password. With our carefully designed password scheme and multi-task learning objective, we achieve both anonymization and deanonymization using the same single network. Extensive experiments show that our method enables multimodal password conditioned anonymizations and deanonymizations, with superior privacy protection compared to existing anonymization methods."

Unpaired image-to-image translation aims to translate images from the source class to target one by providing sufficient data for these classes. Current few-shot translation methods use multiple reference images to describe the target domain through extracting common features. In this paper, we focus on a more specific identity transfer problem and advocate that particular property in each individual image can also benefit generation. We accordingly propose a new multi-reference identity transfer framework by simultaneously making use of particularity and commonality of reference. It is achieved via a semantic pyramid alignment module to make proper use of geometric information for individual images, as well as an attention module to aggregate for the final transformation. Extensive experiments demonstrate the effectiveness of our framework given the promising results in a number of identity transfer applications."

This paper discusses unsupervised domain adaptation (UDA) with target shift, i.e., UDA with the non-identical label distributions of the source and target domains. In practice, this is an important problem; as we do not know labels in target domain datasets, we do not know whether or not its distribution is identical to that in the source domain dataset. Despite the inaccessibility to the shape of label distribution in the target domain, a common approach of modern UDA methods reduces the gap between the feature distributions of source and target domains, which implicitly assumes that the label distributions are identical, resulting in an unsatisfactory performance upon target shift. To overcome this problem, the proposed method, partially shared variational autoencoders (PS-VAEs), uses a pair-wise feature alignment instead of feature distribution matching. PS-VAEs inter-convert domain of each sample by a CycleGAN-based architecture while preserving its label-related content by sharing weights of two domain conversion branches as much as possible. To evaluate the performance of PS-VAEs, we carried out two experiments: UDA from synthesized data to real observation in human-pose estimation (regression) and UDA under controlled target shift intensities with digits datasets (classification). The proposed method outperformed the other methods in the regression task with a large margin while presenting its robustness with various levels of target shift."

Few-shot semantic segmentation aims to learn to segment new object classes with only a few annotated examples, which has a wide range of real-world applications. Most existing methods either focus on the restrictive setting of one-way few-shot segmentation or suffer from incomplete coverage of object regions. In this paper, we propose a novel few-shot semantic segmentation framework based on the prototype representation. Our key idea is to decompose the holistic class representation into a set of part-aware prototypes, capable of capturing diverse and fine-grained object features. In addition, we propose to leverage unlabeled data to enrich our part-aware prototypes, resulting in better modeling of intra-class variations of semantic objects. We develop a novel graph neural network model to generate and enhance the proposed part-aware prototypes based on labeled and unlabeled images. Extensive experimental evaluations on two benchmarks show that our method outperforms the prior art with a sizable margin."

We propose a novel, end-to-end trainable, deep network called ParSeNet that decomposes a 3D point cloud into parametric surface patches, including B-spline patches as well as basic geometric primitives. ParSeNet is trained on a large-scale dataset of man-made 3D shapes and captures high-level semantic priors for shape decomposition. It handles a much richer class of primitives than prior work, and allows us to represent surfaces with higher fidelity. It also produces repeatable and robust parametrizations of a surface compared to purely geometric approaches. We present extensive experiments to validate our approach against analytical and learning-based alternatives. Our source code is publicly available at: https://hippogriff.github.io/parsenet."

Weakly Supervised Object Localization (WSOL) methods only require image level labels as opposed to expensive bounding box annotations required by fully supervised algorithms. We study the problem of learning localization model on target classes with weakly supervised image labels, helped by a fully annotated source dataset. Typically, a WSOL model is first trained to predict class generic objectness scores on an off-the-shelf fully supervised source dataset and then it is progressively adapted to learn the objects in the weakly supervised target dataset. In this work, we argue that learning only an objectness function is a weak form of knowledge transfer and propose to learn a classwise pairwise similarity function that directly compares two input proposals as well. The combined localization model and the estimated object annotations are jointly learned in an alternating optimization paradigm as is typically done in standard WSOL methods. In contrast to the existing work that learns pairwise similarities, our approach optimizes a unified objective with convergence guarantee and it is computationally efficient for large-scale applications. Experiments on the COCO and ILSVRC 2013 detection datasets show that the performance of the localization model improves significantly with the inclusion of pairwise similarity function. For instance, in the ILSVRC dataset, the Correct Localization (CorLoc) performance improves from 72.8% to 78.2% which is a new state-of-the-art for WSOL task in the context of knowledge transfer."

State-of-the-art object detectors exploit multi-branch structure and predict objects at several different scales, although substantially boosted accuracy is acquired, low efficiency is inevitable as fragmented structure is hardware unfriendly. To solve this issue, we propose a packing operator (PackOp) to combine all head branches together at spatial. Packed features are computationally more efficient and allow to use cross-head group normalization (GN) at handy, leading to notable accuracy improvement against the common head-separate GN. All of these are only at the cost of less than 5.7% relative increase on runtime memory and introduction of a few noisy training samples, however, whose side-effects could be diminished by good packing patterns design. With PackOp, we propose a new anchor-free one-stage detector, PackDet, which features a single deeper/longer but narrower head compared to the existing methods: multiple shallow but wide heads. Our best models on COCO test-dev achieve better speed-accuracy balance: 35.1%, 42.3%, 44.0%, 47.4% AP with 22.6, 16.9, 12.4, 4.7 FPS using MobileNet-v2, ResNet-50, ResNet-101, and ResNeXt-101-DCN backbone, respectively. Codes will be released."

In this paper, we propose a novel deep neural network based method, called PUGeo-Net, for upsampling 3D point clouds. PUGeo-Net incorporates discrete differential geometry into deep learning elegantly by learning the first and second fundamental forms that are able to fully represent the local geometry unique up to rigid motion. Specifically, we encode the first fundamental form in a 3 $ imes$ 3 linear transformation matrix $\mathbf{T}$ for each input point. Such a matrix approximates the augmented Jacobian matrix of a local parameterization that encodes the intrinsic information and builds a one-to-one correspondence between the 2D parametric domain and the 3D tangent plane, so that we can lift the adaptively distributed 2D samples learned from the input to 3D space. After that, we use the learned second fundamental form to compute a normal displacement for each generated sample and project it to the curved surface. As a by-product, PUGeo-Net can compute normals for the original and generated points, which is highly desired for surface reconstruction algorithms. We evaluate PUGeo-Net on a wide range of 3D models with sharp features and rich geometric details and observe that PUGeo-Net consistently outperforms state-of-the-art methods in terms of both accuracy and efficiency for upsampling factor 4 - 16. We also verify the geometry-centric nature of PUGeo-Net quantitatively. In addition, PUGeo-Net can handle noisy and non-uniformly distributed inputs well, validating its robustness. The code is publicly available at https://github.com/ninaqy/PUGeo."

Generative 3D shape modeling is a fundamental research area in computer vision and interactive computer graphics, with many real-world applications. This paper investigates the novel problem of generating a 3D point cloud geometry for a shape from a symbolic part tree representation. In order to learn such a conditional shape generation procedure in an end-to-end fashion, we propose a conditional GAN “part tree”-to-“point cloud” model (PT2PC) that disentangles the structural and geometric factors. The proposed model incorporates the part tree condition into the architecture design by passing messages top-down and bottom-up along the part tree hierarchy. Experimental results and user study demonstrate the strengths of our method in generating perceptually plausible and diverse 3D point clouds, given the part tree condition. We also propose a novel structural measure for evaluating if the generated shape point clouds satisfy the part tree conditions."

Despite their strong modeling capacities, Convolutional Neural Networks (CNNs) are often scale-sensitive. For enhancing the robustness of CNNs to scale variance, multi-scale feature fusion from different layers or filters attracts great attention among existing solutions, while the more granular kernel space is overlooked. We bridge this regret by exploiting multi-scale features in a finer granularity. The proposed convolution operation, named Poly-Scale Convolution (PSConv), mixes up a spectrum of dilation rates and tactfully allocates them in the individual convolutional kernels of each filter regarding a single convolutional layer. Specifically, dilation rates vary cyclically along the axes of input and output channels of the filters, aggregating features over a wide range of scales in a neat style. PSConv could be a drop-in replacement of the vanilla convolution in many prevailing CNN backbones, allowing better representation learning without introducing additional parameters and computational complexities. Comprehensive experiments on the ImageNet and MS COCO benchmarks validate the superior performance of PSConv. Code and models are available at https://github.com/d-li14/PSConv"

4-bit and lower precision mobile models are required due to the ever-increasing demand for better energy efficiency in mobile devices. In this work, we report that the activation instability induced by weight quantization (AIWQ) is the key obstacle to sub-4-bit quantization of mobile networks. To alleviate the AIWQ problem, we propose a novel training method called PROgressive-Freezing Iterative Training (PROFIT), which attempts to freeze layers whose weights are affected by the instability problem stronger than the other layers. We also propose a differentiable and unified quantization method (DuQ) and a negative padding idea to support asymmetric activation functions such as h-swish. We evaluate the proposed methods by quantizing MobileNet-v1, v2, and v3 on ImageNet and report that 4-bit quantization offers comparable (within 1.48 % top-1 accuracy) accuracy to full precision baseline. In the ablation study of the 3-bit quantization of MobileNet-v3, our proposed method outperforms the state-of-the-art method by a large margin, 12.86 % of top-1 accuracy. The quantized model and source code is available at https://github.com/EunhyeokPark/PROFIT."

Lifelong learning has attracted much attention, but existing works still struggle to fight catastrophic forgetting and accumulate knowledge over long stretches of incremental learning. In this work, we propose PODNet, a model inspired by representation learning. By carefully balancing the compromise between remembering the old classes and learning new ones, PODNet fights catastrophic forgetting, even over very long runs of small incremental tasks, a setting so far unexplored by current works. PODNet innovates on existing art with an efficient spatial-based distillation-loss applied throughout the model and a representation comprising multiple proxy vectors for each class. We validate those innovations thoroughly, comparing PODNet with three state-of-the-art models on three datasets: CIFAR100, ImageNet100, and ImageNet1000. Our results showcase a significant advantage of PODNet over existing art, with accuracy gains of 12.10, 6.51, and 2.85 percentage points, respectively."

We present a complete classification of minimal problems for generic arrangements of points and lines in space observed partially by three calibrated perspective cameras when each line is incident to at most one point. This is a large class of interesting minimal problems that allows for missing observations in images due to occlusions and missed detections. There is an infinite number of such minimal problems; however, we show that they can be reduced to 140616 equivalence classes by removing superfluous features and relabeling the cameras. We also introduce camera-minimal problems, which are practical for designing minimal solvers, and show how to pick a simplest camera-minimal problem for each minimal problem. This simplification results in 74575 equivalence classes. Only 76 of these were known; the rest are new. In order to identify problems that have potential for practical solving of image matching and 3D reconstruction, we present several smaller natural subfamilies of camera-minimal problems as well as compute solution counts for all camera-minimal problems which have less than 300 solutions for generic data."

Object detection using an oriented bounding box (OBB) can better target rotated objects by reducing the overlap with background areas. Existing OBB approaches are mostly built on horizontal bounding box detectors by introducing an additional angle dimension optimized by a distance loss. However, as the distance loss only minimizes the angle error of the OBB and that it loosely correlates to the IoU, it is insensitive to objects with high aspect ratios. Therefore, a novel loss, Pixels-IoU (PIoU) Loss, is formulated to exploit both the angle and IoU for accurate OBB regression. The PIoU loss is derived from IoU metric with a pixel-wise form, which is simple and suitable for both horizontal and oriented bounding box. To demonstrate its effectiveness, we evaluate the PIoU loss on both anchor-based and anchor-free frameworks. The experimental results show that PIoU loss can dramatically improve the performance of OBB detectors, particularly on objects with high aspect ratios and complex backgrounds. Besides, previous evaluation datasets did not include scenarios where the objects have high aspect ratios, hence a new dataset, Retail50K, is introduced to encourage the community to adapt OBB detectors for more complex environments."

Image quality assessment (IQA) is the key factor for the fast development of image restoration (IR) algorithms. The most recent IR methods based on Generative Adversarial Networks (GANs) have achieved significant improvement in visual performance, but also presented great challenges for quantitative evaluation. Notably, we observe an increasing inconsistency between perceptual quality and the evaluation results. Then we raise two questions: (1) Can existing IQA methods objectively evaluate recent IR algorithms? (2) When focus on beating current benchmarks, are we getting better IR algorithms? To answer these questions and promote the development of IQA methods, we contribute a large-scale IQA dataset, called Perceptual Image Processing Algorithms (PIPAL) dataset. Especially, this dataset includes the results of GAN-based methods, which are missing in previous datasets. We collect more than 1.13 million human judgments to assign subjective scores for PIPAL images using the more reliable ``Elo system''. Based on PIPAL, we present new benchmarks for both IQA and super-resolution methods. Our results indicate that existing IQA methods cannot fairly evaluate GAN-based IR algorithms. While using appropriate evaluation methods is important, IQA methods should also be updated along with the development of IR algorithms. At last, we improve the performance of IQA networks on GAN-based distortions by introducing anti-aliasing pooling. Experiments show the effectiveness of the proposed method."

Siamese neural network has been well investigated by tracking frameworks due to its fast speed and high accuracy. However, very few efforts were spent on background-extraction by those approaches. In this paper, a Pixel to Global Matching Network (PG-Net) is proposed to suppress the influence of background in search image while achieving state-of-the-art tracking performance. To achieve this purpose, each pixel on search feature is utilized to calculate the similarity with global template feature. This calculation method can appropriately reduce the matching area, thus introducing less background interference. In addition, we propose a new tracking framework to perform correlation-shared tracking and multiple losses for training, which not only reduce the computational burden but also improve the performance. We conduct comparison experiments on various public tracking datasets, which obtains state-of-the-art performance while running with fast speed."

Deep convolutional neural networks (DCNNs) have shown dominant performance in the task of super-resolution (SR). However, their heavy memory cost and computation overhead significantly restrict their practical deployments on resource-limited devices, which mainly arise from the floating-point storage and operations between weights and activations. Although previous endeavors mainly resort to fixed-point operations, quantizing both weights and activations with fixed coding lengths may cause significant performance drop, especially on low bits. Specifically, most state-of-the-art SR models without batch normalization have a large dynamic quantization range, which also serves as another cause of performance drop. To address these two issues, we propose a new quantization scheme termed PArameterized Max Scale (PAMS), which applies the trainable truncated parameter to explore the upper bound of the quantization range adaptively. Finally, a structured knowledge transfer (SKT) loss is introduced to fine-tune the quantized network. Extensive experiments demonstrate that the proposed PAMS scheme can well compress and accelerate the existing SR models such as EDSR and RDN. Notably, 8-bit PAMS-EDSR improves PSNR on Set5 benchmark from 32.095dB to 32.124dB with 2.42$ imes$ compression ratio, which achieves a new state-of-the-art."

Large-scale pre-training methods of learning cross-modal representations on image-text pairs are becoming popular for vision-language tasks. While existing methods simply concatenate image region features and text features as input to the model to be pre-trained and use self-attention to learn image-text semantic alignments in a brute force manner, in this paper, we propose a new learning method Oscar, which uses object tags detected in images as anchor points to significantly ease the learning of alignments. Our method is motivated by the observation that the salient objects in an image can be accurately detected, and are often mentioned in the paired text. We pre-train an Oscar model on the public corpus of 6.5 million text-image pairs, and fine-tune it on downstream tasks, creating new state-of-the-arts on six well-established vision-language understanding and generation tasks."