ECCV 2020 Papers

Benefiting from deep learning research and large-scale datasets, saliency prediction has achieved significant success in the past decade. However, it still remains challenging to predict saliency maps on images in new domains that lack sufficient data for data-hungry models. To solve this problem, we propose a few-shot transfer learning paradigm for saliency prediction, which enables efficient transfer of knowledge learned from the existing large-scale saliency datasets to a target domain with limited labeled examples. Specifically, very few target domain examples are used as the reference to train a model with a source domain dataset such that the training process can converge to a local minimum in favor of the target domain. Then, the learned model is further fine-tuned with the reference. The proposed framework is gradient-based and model-agnostic. We conduct comprehensive experiments and ablation study on various source domain and target domain pairs. The results show that the proposed framework achieves a significant performance improvement. The code is publicly available at https://github.com/luoyan407/n-reference."

We propose an interpretable Capsule Network, iCaps, for image classification. A capsule is a group of neurons nested inside each layer, and the one in the last layer is called a class capsule, which is a vector whose norm indicates a predicted probability for the class. Using the class capsule, existing Capsule Networks already provide some level of interpretability. However, there are two limitations which degrade its interpretability: 1) the class capsule also includes classification-irrelevant information, and 2) entities represented by the class capsule overlap. In this work, we address these two limitations using a novel class-supervised disentanglement algorithm and an additional regularizer, respectively. Through quantitative and qualitative evaluations on three datasets, we demonstrate that the resulting classifier, iCaps, provides a prediction along with clear rationales behind it with no performance degradation."

In the past few years, we have witnessed the great progress of image super-resolution (SR) thanks to the power of deep learning. However, a major limitation of the current image SR approaches is that they assume a pre-determined degradation model or kernel, e.g. bicubic, controls the image degradation process. This makes them easily fail to generalize in a real-world or non-ideal environment since the degradation model of an unseen image may not obey the pre-determined kernel used when training the SR model. In this work, we present a simple yet effective zero-shot image super-resolution model. Our zero-shot SR model learns an image-specific super-resolution network (SRN) from a low-resolution input image alone, without relying on external training sets. To circumvent the difficulty caused by the unknown internal degradation model of an image, we propose to learn an image-specific degradation simulation network (DSN) together with our image-specific SRN. Specifically, we exploit the depth information, naturally indicating the scales of local image patches, of an image to extract the unpaired high/low-resolution patch collection to train our networks. According to the benchmark test on four datasets with depth labels or estimated depth maps, our proposed depth guided degradation model learning based image super-resolution (DGDML-SR) achieves visually pleasing results and can outperform the state-of-the-arts in perceptual metrics."

We present a novel approach to geolocalising panoramic images on a 2-D cartographic map based on learning a low dimensional embedded space, which allows a comparison between an image captured at a location and local neighbourhoods of the map. The representation is not sufficiently discriminatory to allow localisation from a single image, but when concatenated along a route, localisation converges quickly, with over 90% accuracy being achieved for routes of around 200m in length when using Google Street View and Open Street Map data. The method generalises a previous fixed semantic feature based approach and achieves significantly higher localisation accuracy and faster convergence."

The transferability of adversarial examples across deep neural network (DNN) models is the crux of a spectrum of black-box attacks. In this paper, we propose a novel method to enhance the black-box transferability of baseline adversarial examples. By establishing a linear mapping of the intermediate-level discrepancies (between a set of adversarial inputs and their benign counterparts) for predicting their evoked adversarial loss, we manage to take full advantage of the optimization procedure of baseline attacks. We conduct extensive experiments to verify the effectiveness of our method on CIFAR-100 and ImageNet. Experimental results demonstrate that it outperforms previous state-of-the-arts by large margins. Our code will be made publicly available."

Generating new visual tasks requires additional datasets, and it costs a considerable effort. We propose a new method of domain adaptation for merging multiple models with less effort than creating an additional dataset. This method merges pre-trained models in different domains using the glue layers and the generative model, which feeds latent features to train the glue layers without an additional dataset. We also propose the generative model created by knowledge distillation from pre-trained models. It also allows reusing the dataset to create latent features for training the glue layers. We apply this method to object detection in a low-light situation. The “YOLO in the Dark” contains two models, “Learning to See in the Dark” and YOLO. We report the new method and the result of domain adaptation that detect objects from raw short-exposure low-light images. The “YOLO in the Dark” costs fewer computing resources compared to the naive approach."

We propose a novel Generative Adversarial Network (XingGAN or CrossingGAN) for person image generation tasks, i.e., translating the pose of a given person to a desired one. The proposed Xing generator consists of two generation branches that model the person's appearance and shape information, respectively. Moreover, we propose two novel blocks to effectively transfer and update the person's shape and appearance embeddings in a crossing way to mutually improve each other, which has not been considered by any other existing GAN-based image generation work. Extensive experiments on two challenging datasets, i.e., Market-1501 and DeepFashion, demonstrate that the proposed XingGAN advances the state-of-the-art performance both in terms of objective quantitative scores and subjective visual realness. The source code and trained models are available at https://github.com/Ha0Tang/XingGAN."

Video-to-video synthesis is a powerful tool for converting high-level semantic inputs to photorealistic videos. However, while existing vid2vid methods can maintain short-term temporal consistency, they fail to ensure long-term consistency in the outputs. This is because they generate each frame only based on the past few frames. They lack knowledge of the 3D world being generated. In this work, we propose a framework for utilizing all past generated frames when synthesizing each frame. This is achieved by condensing the 3D world generated so far into a physically-grounded estimate of the current frame, which we call the guidance image. A novel module is also proposed to take advantage of the information stored in the guidance images. Extensive experimental results on several challenging datasets verify the effectiveness of our method in achieving world consistency - the output video is consistent within the entire generated 3D world.

Explaining a deep learning model can help users understand its behavior and allow researchers to discern its shortcomings. Recent work has primarily focused on explaining models for tasks like image classification or visual question answering. In this paper, we introduce Salient Attributes for Network Explanation (SANE) to explain image similarity models, where a model's output is a score measuring the similarity of two inputs rather than a classification score. In this task, an explanation depends on both of the input images, so standard methods do not apply. Our SANE explanations pairs a saliency map identifying important image regions with an attribute that best explains the match. We find that our explanations provide additional information not typically captured by saliency maps alone, and can also improve performance on the classic task of attribute recognition. Our approach's ability to generalize is demonstrated on two datasets from diverse domains, Polyvore Outfits and Animals with Attributes 2."

Recently, intermediate feature maps of pre-trained convolutional neural networks have shown significant perceptual quality improvements, when they are used in the loss function for training new networks. It is believed that these features are better at encoding the perceptual quality and provide more efficient representations of input images compared to other perceptual metrics such as SSIM and PSNR. However, there have been no systematic studies to determine the underlying reason. Due to the lack of such an analysis, it is not possible to evaluate the performance of a particular set of features or to improve the perceptual quality even more by carefully selecting a subset of features from a pre-trained CNN. This work shows that the capabilities of pre-trained deep CNN features in optimizing the perceptual quality are correlated with their success in capturing basic human visual perception characteristics. In particular, we focus our analysis on fundamental aspects of human perception, such as the contrast sensitivity and orientation selectivity. We introduce two new formulations to measure the frequency and orientation selectivity of the features learned by convolutional layers for evaluating deep features learned by widely-used deep CNNs such as VGG-16. We demonstrate that the pre-trained CNN features which receive higher scores are better at predicting human quality judgment. Furthermore, we show the possibility of using our method to select deep features to form a new loss function, which improves the image reconstruction quality for the well-known single-image super-resolution problem. "

This paper investigates the task of 2D human whole-body pose estimation, which aims to localize dense landmarks on the entire human body including face, hands, body, and feet. As existing datasets do not have whole-body annotations, previous methods have to assemble different deep models trained independently on different datasets of the human face, hand, and body, struggling with dataset biases and large model complexity. To fill in this blank, we introduce COCO-WholeBody which extends COCO dataset with whole-body annotations. To our best knowledge, it is the first benchmark that has manual annotations on the entire human body, including 133 dense landmarks with 68 on the face, 42 on hands and 23 on the body and feet. A single-network model, named ZoomNet, is devised to take into account the hierarchical structure of the full human body to solve the scale variation of different body parts of the same person. ZoomNet is able to significantly outperform existing methods on the proposed COCO-WholeBody dataset. Extensive experiments show that COCO-WholeBody not only can be used to train deep models from scratch for whole-body pose estimation but also can serve as a powerful pre-training dataset for many different tasks such as facial landmark detection and hand keypoint estimation. The dataset is publicly available at \url{https://github.com/jin-s13/COCO-WholeBody}."

We introduce an automatic, end-to-end method for recovering the 3D pose and shape of dogs from monocular internet images. The large variation in shape between dog breeds, and significant occlusion and low quality of internet images makes this a challenging problem. We learn a richer prior over shapes than previous work, which helps regularize parameter estimation. We demonstrate results on the Stanford Dog Dataset, an ``in-the-wild'' dataset of 20,580 dog images for which we have collected 2D joint and silhouette annotations to split for training and evaluation. In order to capture the large shape variety of dogs, we show that the natural variation in the 2D dataset is enough to learn a detailed 3D prior through expectation maximisation (EM). As a by-product of training, we generate a new parameterised model (including limb scaling) SMBLD which we release alongside the annotation dataset to the research community."

In this work we address the problem of autonomous 3D exploration of an unknown indoor environment using a depth camera. We cast the problem as the estimation of the Next Best View (NBV) that maximises the coverage of the unknown area. We do this by re-formulating NBV estimation as a classification problem and we propose a novel learning-based metric that encodes both, the current 3D observation (a depth frame) and the history of the ongoing reconstruction. One of the major contributions of this work is about introducing a new representation for the 3D reconstruction history as an auxiliary utility map which is efficiently coupled with the current depth observation. With both pieces of information, we train a light-weight CNN, named ExHistCNN, that estimates the NBV as a set of directions towards which the depth sensor finds most unexplored areas. We perform extensive evaluation on both synthetic and real room scans demonstrating that the proposed ExHistCNN is able to approach the exploration performance of an oracle using the complete knowledge of the 3D environment."

We investigate the role of self-supervised learning (SSL) in the context of few-shot learning. Although recent research has shown the benefits of SSL on large unlabeled datasets, its utility on small datasets is relatively unexplored. We find that SSL reduces the relative error rate of few-shot meta-learners by 4%-27%, even when the datasets are small and only utilizing images within the datasets. The improvements are greater when the training set is smaller or the task is more challenging. Although the benefits of SSL may increase with larger training sets, we observe that SSL can hurt the performance when the distributions of images used for meta-learning and SSL are different. We conduct a systematic study by varying the degree of domain shift and analyzing the performance of several meta-learners on a multitude of domains. Based on this analysis we present a technique that automatically selects images for SSL from a large, generic pool of unlabeled images for a given dataset that provides further improvements."

The quality of image generation and manipulation is reaching impressive levels, making it exceedingly difficult for a human to distinguish between what is real and what is fake. However, deep networks can still pick up on the subtle artifacts in these doctored images. We seek to understand what properties of these fake images make them detectable and identify what generalizes across different model architectures, datasets, and variations in training. We use a patch-based classifier with limited receptive fields to focus on low-level artifacts rather than global semantics, and use patch-wise predictions to localize the manipulated regions. We further show a technique to exaggerate these detectable properties and demonstrate that even when the image generator is adversarially finetuned against the fakeness classifier, it is still imperfect and makes detectable mistakes in similar regions of the image."

This paper targets at discovering what a deep uncalibrated photometric stereo network learns to resolve the problem’s inherent ambiguity, and designing an effective network architecture based on the new insight to improve the performance. The recently proposed deep uncalibrated photometric stereo method achieved promising results in estimating directional lightings. However, what specifically inside the network contributes to its success remains a mystery. In this paper, we analyze the features learned by this method and find that they strikingly resemble attached shadows, shadings, and specular highlights, which are known to provide useful clues in resolving the generalized bas-relief (GBR) ambiguity. Based on this insight, we propose a guided calibration network, named GCNet, that explicitly leverages object shape and shading information for improved lighting estimation. Experiments on synthetic and real datasets show that GCNet achieves improved results in lighting estimation for photometric stereo, which echoes the findings of our analysis. We further demonstrate that GCNet can be directly integrated with existing calibrated methods to achieve improved results on surface normal estimation."

We systematically compare and analyze a set of key components in unsupervised optical flow to identify which photometric loss, occlusion handling, and smoothness regularization is most effective. Alongside this investigation we construct a number of novel improvements to unsupervised flow models, such as cost volume normalization, stopping the gradient at the occlusion mask, encouraging smoothness before upsampling the flow field, and continual self-supervision with image resizing. By combining the results of our investigation with our improved model components, we are able to present a new unsupervised flow technique that significantly outperforms the previous unsupervised state-of-the-art and performs on par with supervised FlowNet2 on the KITTI 2015 dataset, while also being significantly simpler than related approaches."

We present a conceptually simple, flexible and effective framework for weight generating networks. Our approach is general that unifies two current distinct and extremely effective SENet and CondConv into the same framework on weight space. The method, called WeightNet, generalizes the two methods by simply adding one more grouped fully-connected layer to the attention activation layer. We use the WeightNet, composed entirely of (grouped) fully-connected layers, to directly output the convolutional weight. WeightNet is easy and memory-conserving to train, on the kernel space instead of the feature space. Because of the flexibility, our method outperforms existing approaches on both ImageNet and COCO detection tasks, achieving better Accuracy-FLOPs and Accuracy-Parameter trade-offs. The framework on the flexible weight space has the potential to further improve the performance."

We propose novel approaches for simultaneously identifying important weights of a convolutional neural network (ConvNet) and providing more attention to the important weights during training. More formally, we identify two characteristics of a weight, its magnitude and its location, which can be linked with the importance of the weight. By targeting these characteristics of a weight during training, we develop two separate weight excitation (WE) mechanisms via weight reparameterization-based backpropagation modifications. We demonstrate significant improvements over popular baseline ConvNets on multiple computer vision applications using WE (e.g. 1.3% accuracy improvement over ResNet50 baseline on ImageNet image classification, etc.). These improvements come at no extra computational cost or ConvNet structural change during inference. Additionally, including WE methods in a convolution block is straightforward, requiring few lines of extra code. Lastly, WE mechanisms can provide complementary benefits when used with external attention mechanisms such as the popular Squeeze-and-Excitation attention block."