ECCV 2020 论文列表

Generating a new font library is a very labor-intensive and time-consuming job for glyph-rich scripts. Despite the remarkable success of existing font generation methods, they have significant drawbacks; they require a large number of reference images to generate a new font set, or they fail to capture detailed styles with only a few samples. In this paper, we focus on compositional scripts, a widely used letter system in the world, where each glyph can be decomposed by several components. By utilizing the compositionality of compositional scripts, we propose a novel font generation framework, named Dual Memory-augmented Font Generation Network (DM-Font), which enables us to generate a high-quality font library with only a few samples. We employ memory components and global-context awareness in the generator to take advantage of the compositionality. In the experiments on Korean-handwriting fonts and Thai-printing fonts, we observe that our method generates a significantly better quality of samples with faithful stylization compared to the state-of-the-art generation methods quantitatively and qualitatively. Source code is available at https://github.com/clovaai/dmfont."

Many few-shot learning models focus on recognising images. In contrast, we tackle a challenging task of few-shot action recognition from videos. We build on a C3D encoder for spatio-temporal video blocks to capture short-range action patterns. Such encoded blocks are aggregated by permutation-invariant pooling to make our approach robust to varying action lengths and long-range temporal dependencies whose patterns are unlikely to repeat even in clips of the same class. Subsequently, the pooled representations are combined into simple relation descriptors which encode so-called query and support clips. Finally, relation descriptors are fed to the comparator with the goal of similarity learning between query and support clips. Importantly, to re-weight block contributions during pooling, we exploit spatial and temporal attention modules and self-supervision. In naturalistic clips (of the same class) there exists a temporal distribution shift--the locations of discriminative temporal action hotspots vary. Thus, we permute blocks of a clip and align the resulting attention regions with similarly permuted attention regions of non-permuted clip to train the attention mechanism invariant to block (and thus long-term hotspot) permutations. Our method outperforms the state of the art on the HMDB51, UCF101, miniMIT datasets."

The impressive performance of deep convolutional neural networks in single-view 3D reconstruction suggests that these models perform non-trivial reasoning about the 3D structure of the output space. However, recent work has challenged this belief, showing that complex encoder-decoder architectures perform similarly to nearest-neighbor baselines or simple linear decoder models that exploit large amounts of per category data in standard benchmarks. On the other hand settings where 3D shape must be inferred for new categories with few examples are more natural and require models that generalize about shapes. In this work we demonstrate experimentally that naive baselines do not apply when the goal is to learn to reconstruct novel objects using very few examples, and that in a mph{few-shot} learning setting, the network must learn concepts that can be applied to new categories, avoiding rote memorization. To address deficiencies in existing approaches to this problem, we propose three approaches that efficiently integrate a class prior into a 3D reconstruction model, allowing to account for intra-class variability and imposing an implicit compositional structure that the model should learn. Experiments on the popular ShapeNet database demonstrate that our method significantly outperform existing baselines on this task in the few-shot setting."

Few-shot segmentation has recently generated great popularity, addressing a challenging yet important problem of segmenting objects from unseen categories with scarce annotated support images. The crux of few-shot segmentation is to extract object information from the support image and then propagate it to guide the segmentation of query images. In this paper, we propose the Democratic Attention Network (DAN) for few-shot semantic segmentation. We introduce the democratized graph attention mechanism, which can activate more pixels on the object to establish a robust correspondence between support and query images. Thus, the network is able to propagate more guiding information of foreground objects from support to query images, enhancing its robustness and generalizability to new objects. Furthermore, we propose multi-scale guidance by designing a refinement fusion unit to fuse features from intermediate layers for the segmentation of the query image. This offers an efficient way of leveraging multi-level semantic information to achieve more accurate segmentation. Extensive experiments on three benchmarks demonstrate that the proposed DAN achieves the new state-of-the-art performance, surpassing the previous methods by large margins. The thorough ablation studies further reveal its great effectiveness for few-shot semantic segmentation."

We address the problem of anomaly detection in videos. The goal is to identify unusual behaviours automatically by learning exclusively from normal videos. Most existing approaches are usually data-hungry and have limited generalization abilities. They usually need to be trained on a large number of videos from a target scene to achieve good results in that scene. In this paper, we propose a novel few-shot scene-adaptive anomaly detection problem to address the limitations of previous approaches. Our goal is to learn to detect anomalies in a previously unseen scene with only a few frames. A reliable solution for this new problem will have huge potential in real-world applications since it is expensive to collect a massive amount of data for each target scene. We propose a meta-learning based approach for solving this new problem; extensive experimental results demonstrate the effectiveness of our proposed method. "

Detecting objects and estimating their viewpoint in images are key tasks of 3D scene understanding. Recent approaches have achieved excellent results on very large benchmarks for object detection and viewpoint estimation. However, performances are still lagging behind for novel object categories with few samples. In this paper, we tackle the problems of few-shot object detection and few-shot viewpoint estimation. We propose a meta-learning framework that can be applied to both tasks, possibly including 3D data. Our models improve the results on objects of novel classes by leveraging on rich feature information originating from base classes with many samples. A simple joint feature embedding module is proposed to make the most of this feature sharing. Despite its simplicity, our method outperforms state-of-the-art methods by a large margin on a range of datasets, including PASCAL VOC and MS COCO for few-shot object detection, and Pascal3D+ and ObjectNet3D for few-shot viewpoint estimation. And for the first time, we tackle the combination of both few-shot tasks, on ObjectNet3D, showing promising results."

Federated Learning enables visual models to be trained on-device, bringing advantages for user privacy (data need never leave the device), but challenges in terms of data diversity and quality. Whilst typical models in the datacenter are trained using data that are independent and identically distributed (IID), data at source are typically far from IID. Furthermore, differing quantities of data are typically available at each device (imbalance). In this work, we characterize the effect these real-world data distributions have on distributed learning, using as a benchmark the standard Federated Averaging (FedAvg) algorithm. To do so, we introduce two new large-scale datasets for species and landmark classification, with realistic per-user data splits that simulate real-world edge learning scenarios. We also develop two new algorithms (FedVC, FedIR) that intelligently resample and reweight over the client pool, bringing large improvements in accuracy and stability in training. The datasets are made available online."

Photometric loss is widely used for self-supervised depth and egomotion estimation. However, the loss landscapes induced by photometric differences are often problematic for optimization, caused by plateau landscapes for pixels in texture-less regions or multiple local minima for less discriminative pixels. In this work, feature-metric loss is proposed and defined on feature representation, where the feature representation is also learned in a self-supervised manner and regularized by both first-order and second-order derivatives to constrain the loss landscapes to form proper convergence basins. Comprehensive experiments and detailed analysis via visualization demonstrate the effectiveness of the proposed feature-metric loss. In particular, our method improves state-of-the-art methods on KITTI from 0.885 to 0.925 measured by $\delta_1$ for depth estimation, and significantly outperforms previous method for visual odometry."

Real-world data often follow a long-tailed distribution as the frequency of each class is typically different. For example, a dataset can have a large number of under-represented classes and a few classes with more than sufficient data. However, a model to represent the dataset is usually expected to have reasonably homogeneous performances across classes. Introducing class-balanced loss and advanced methods on data re-sampling and augmentation are among the best practices to alleviate the data imbalance problem. However, the other part of the problem about the under-represented classes will have to rely on additional knowledge to recover the missing information.

In this paper, we are aiming for a general reference-based super-resolution setting: it does not require the low-resolution image and the high-resolution reference image to be well aligned or with a similar texture. Instead, we only intend to transfer the relevant textures from reference images to the output super-resolution image. To this end, we engaged neural texture transfer to swap texture features between the low-resolution image and the high-resolution reference image. We identify the importance of designing a super-resolution task-specific features rather than classification oriented features for neural texture transfer, making the feature extractor more compatible with the image synthesis task. We develop an end-to-end training framework for the reference-based super-resolution task, where the feature encoding network prior to matching and swapping is jointly trained with the image synthesis network. We also discover that learning the high-frequency residual is an effective way for the reference-based super-resolution task. Without bells and whistles, the proposed method E2ENT2 achieved better performance than state-of-the method (i.e., SRNTT with five loss functions) with only two basic loss functions. Extensive experimental results on several datasets demonstrate that the proposed method E2ENT2 can achieve superior performance to existing best models both quantitatively and qualitatively."

Feature interactions across space and scales underpin modern visual recognition systems because they introduce beneficial visual contexts. Conventionally, spatial contexts are passively hidden in the CNN's increasing receptive fields or actively encoded by non-local convolution. Yet, the non-local spatial interactions are not across scales, and thus they fail to capture the non-local contexts of objects (or parts) residing in different scales. To this end, we propose a fully active feature interaction across both space and scales, called Feature Pyramid Transformer (FPT). It transforms any feature pyramid into another feature pyramid of the same size but with richer contexts, by using three specially designed transformers in self-level, top-down, and bottom-up interaction fashion. FPT serves as a generic visual backbone with fair computational overhead. We conduct extensive experiments in both instance-level (i.e., object detection and instance segmentation) and pixel-level segmentation tasks, using various backbones and head networks, and observe consistent improvement over all the baselines and the state-of-the-art methods."

Knowledge Distillation (KD) transfers the knowledge from a cumbersome teacher model to a lightweight student network. Since a single image may reasonably relate to several categories, the one-hot label would inevitably introduce the encoding noise. From this perspective, we systematically analyze the distillation mechanism and demonstrate that the L2-norm of the feature in penultimate layer would be too large under the influence of label noise, and the temperature T in KD could be regarded as a correction factor for L2-norm to suppress the impact of noise. Noticing different samples suffer from varying intensities of label noise, we further propose a simple yet effective feature normalized knowledge distillation which introduces the sample specific correction factor to replace the unified temperature T for better reducing the impact of noise. Extensive experiments show that the proposed method surpasses standard KD as well as self-distillation significantly on Cifar-100, CUB-200-2011 and Stanford Cars datasets. The codes are in https://github.com/aztc/FNKD"

Recent state-of-the-art semi-supervised learning (SSL) methods use a combination of image-based transformations and consistency regularization as core components. Such methods, however, are limited to simple transformations such as traditional data augmentation or convex combinations of two images. In this paper, we propose a novel learned feature-based refinement and augmentation method that produces a varied set of complex transformations. Importantly, these transformations also use information from both within-class and across-class prototypical representations that we extract through clustering. We use features already computed across iterations by storing them in a memory bank, obviating the need for significant extra computation. These transformations, combined with traditional image-based augmentation, are then used as part of the consistency-based regularization loss. We demonstrate that our method is comparable to current state of art for smaller datasets (CIFAR-10 and SVHN) while being able to scale up to larger datasets such as CIFAR-100 and mini-Imagenet where we achieve significant gains over the state of art ( extit{e.g.,} absolute 17.44\% gain on mini-ImageNet). We further test our method on DomainNet, demonstrating better robustness to out-of-domain unlabeled data, and perform rigorous ablations and analysis to validate the method."

Fast person re-identification (ReID) aims to search person images quickly and accurately. The main idea of recent fast ReID methods is the hashing algorithm, which learns compact binary codes and performs fast Hamming distance and counting sort. However, a very long code is needed for high accuracy (e.g. 2048), which compromises search speed. In this work, we introduce a new solution for fast ReID by formulating a novel Coarse-to-Fine (CtF) hashing code search strategy, which complementarily uses short and long codes, achieving both faster speed and better accuracy. It uses shorter codes to coarsely rank broad matching similarities and longer codes to refine only a few top candidates for more accurate instance ReID. Specifically, we design an All-in-One (AiO) framework together with a Distance Threshold Optimization (DTO) algorithm. In AiO, we simultaneously learn and enhance multiple codes of different lengths in a single model. It learns multiple codes in a pyramid structure, and encourage shorter codes to mimic longer codes by self-distillation. DTO solves a complex threshold search problem by a simple optimization process, and the balance between accuracy and speed is easily controlled by a single parameter. It formulates the optimization target as a $F_{eta}$ score that can be optimised by Gaussian cumulative distribution functions. Experimental results on 2 datasets show that our proposed method (CtF) is not only 8% more accurate but also 5x faster than contemporary hashing ReID methods. Compared with non-hashing ReID methods, CtF is 50x faster with comparable accuracy. Code is available at https://github.com/wangguanan/light-reid."

Data augmentation methods are indispensable heuristics to boost the performance of deep neural networks, especially in image recognition tasks. Recently, several studies have shown that augmentation strategies found by search algorithms outperform hand-made strategies. Such methods employ black-box search algorithms over image transformations with continuous or discrete parameters and require a long time to obtain better strategies. In this paper, we propose a differentiable policy search pipeline for data augmentation, which is much faster than previous methods. We introduce approximate gradients for several transformation operations with discrete parameters as well as a differentiable mechanism for selecting operations. As the objective of training, we minimize the distance between the distributions of augmented and original data, which can be differentiated. We show that our method, Faster AutoAugment, achieves significantly faster searching than prior methods without a performance drop."

We developed a real-time, high-quality semi-supervised video object segmentation algorithm. Its accuracy is on par with the most accurate, time-consuming online-learning model, while its speed is similar to the fastest template-matching method with sub-optimal accuracy. The core component of the model is a novel global context module that effectively summarizes and propagates information through the entire video. Compared to previous approaches that only use one frame or a few frames to guide the segmentation of the current frame, the global context module uses all past frames. Unlike the previous state-of-the-art space-time memory network that caches a memory at each spatio-temporal position, the global context module uses a fixed-size feature representation. Therefore, it uses constant memory regardless of the video length and costs substantially less memory and computation. With the novel module, our model achieves top performance on standard benchmarks at a real-time speed."

We propose a neural rendering-based system that creates head avatars from a single photograph. Our approach models a person's appearance by decomposing it into two layers. The first layer is a pose-dependent coarse image that is synthesized by a small neural network. The second layer is defined by a pose-independent texture image that contains high-frequency details. The texture image is generated offline, warped and added to the coarse image to ensure a high effective resolution of synthesized head views. We compare our system to analogous state-of-the-art systems in terms of visual quality and speed. The experiments show significant inference speedup over previous neural head avatar models for a given visual quality. We also report on a real-time smartphone-based implementation of our system."

Conventional supervised super-resolution (SR) approaches are trained with massive external SR datasets but fail to exploit desirable properties of the given test image.On the other hand, self-supervised SR approaches utilize the internal information within a test image but suffer from computational complexity in run-time. In this work, we observe the opportunity for further improvement of the performance of SISR without changing the architecture of conventional SR networks by practically exploiting additional information given from the input image. In the training stage, we train the network via meta-learning; thus, the network can quickly adapt to any input image at test time. Then, in the test stage, parameters of this meta-learned network are rapidly fine-tuned with only a few iterations by only using the given low-resolution image. The adaptation at the test time takes full advantage of patch-recurrence property observed in natural images. Our method effectively handles unknown SR kernels and can be applied to any existing model. We demonstrate that the proposed model-agnostic approach consistently improves the performance of conventional SR networks on various benchmark SR datasets."

Segmentation, and an Attribute Localization Dataset","In this work, we focus on the task of instance segmentation with attribute localization. This unifies instance segmentation (detect and segment each object instance) and visual categorization of fine-grained attributes (classify one or multiple attributes). The proposed task requires both localizing an object and describing its properties. To illustrate the various aspects of this task, we focus on the domain of fashion and introduce Fashionpedia as a step toward mapping out the visual aspects of the fashion world. Fashionpedia consists of two parts: (1) an ontology built by fashion experts containing 27 main apparel categories, 19 apparel parts, and 294 fine-grained attributes and their relationships and (2) a dataset consisting of everyday and celebrity event fashion images annotated with segmentation masks and their associated fine-grained attributes, built upon the backbone of the Fashionpedia ontology. In order to solve this challenging task, we propose a novel Attribute-Mask R-CNN model to jointly perform instance segmentation and localized attribute recognition, and provide a novel evaluation metric for the task. Fashionpedia is available at https://fashionpedia.github.io/home/. "