ECCV 2018 论文列表

We propose a novel Convolutional Neural Network (CNN) compression algorithm based on coreset representations of filters. We exploit the redundancies extant in the space of CNN weights and neuronal activations (across samples) in order to obtain compression. Our method requires no retraining, is easy to implement, and obtains state-of-the-art compression performance across a wide variety of CNN architectures. Coupled with quantization and Huffman coding, we create networks that provide AlexNet-like accuracy, with a memory footprint that is $832 imes$ smaller than the original AlexNet, while also introducing significant reductions in inference time as well. Additionally these compressed networks when fine-tuned, successfully generalize to other domains as well.

Do convolutional networks really need a fixed feed-forward structure? What if, after identifying the high-level concept of an image, a network could move directly to a layer that can distinguish fine-grained differences? Currently, a network would first need to execute sometimes hundreds of intermediate layers that specialize in unrelated aspects. Ideally, the more a network already knows about an image, the better it should be at deciding which layer to compute next. In this work, we propose convolutional networks with adaptive inference graphs (ConvNet-AIG) that adaptively define their network topology conditioned on the input image. Following a high-level structure similar to residual networks (ResNets), ConvNet-AIG decides for each input image on the fly which layers are needed. In experiments on ImageNet we show that ConvNet-AIG learns distinct inference graphs for different categories. Both ConvNet-AIG with 50 and 101 layers outperform their ResNet counterpart, while using 20% and 33% less computations respectively. By grouping parameters into layers for related classes and only executing relevant layers, ConvNet-AIG improves both efficiency and overall classification quality. Lastly, we also study the effect of adaptive inference graphs on the susceptibility towards adversarial examples. We observe that ConvNet-AIG shows a higher robustness than ResNets, complementing other known defense mechanisms.

ConvNets and Imagenet have driven the recent success of deep learning for image classification. However, the marked slowdown in performance improvement combined with the lack of robustness of neural networks to adversarial examples and their tendency to exhibit undesirable biases question the reliability of these methods. This work investigates these questions from the perspective of the end-user by using human subject studies and explanations. The contribution of this study is threefold. We first experimentally demonstrate that the accuracy and robustness of ConvNets measured on Imagenet are vastly underestimated. Next, we show that explanations can mitigate the impact of misclassified adversarial examples from the perspective of the end-user. We finally introduce a novel tool for uncovering the undesirable biases learned by a model. These contributions also show that explanations are a valuable tool both for improving our understanding of ConvNets' predictions and for designing more reliable models

In recent years, deep Convolutional Neural Networks (CNNs) have broken all records in salient object detection. However, training such a deep model requires a large amount of manual annotations. Our goal is to overcome this limitation by automatically converting an existing deep contour detection model into a salient object detection model without using any manual salient object masks. For this purpose, we have created a deep network architecture, namely Contour-to-Saliency Network (C2S-Net), by grafting a new branch onto a well-trained contour detection network. Therefore, our C2S-Net has two branches for performing two different tasks: 1) predicting contours with the original contour branch, and 2) estimating per-pixel saliency score of each image with the newly-added saliency branch. To bridge the gap between these two tasks, we further propose a contour-to-saliency transferring method to automatically generate salient object masks which can be used to train the saliency branch from outputs of the contour branch. Finally, we introduce a novel alternating training pipeline to gradually update the network parameters. In this scheme, the contour branch generates saliency masks for training the saliency branch, while the saliency branch, in turn, feeds back saliency knowledge in the form of saliency-aware contour labels, for fine-tuning the contour branch. The proposed method achieves state-of-the-art performance on five well-known benchmarks, outperforming existing fully supervised methods while also maintaining high efficiency.

We study the task of image inpainting, which is to fill in the missing region of an incomplete image with plausible contents. To this end, we propose a learning-based approach to generate visually coherent completion given a high-resolution image with missing components. In order to overcome the difficulty to directly learn the distribution of high-dimensional image data, we divide the task into inference and translation as two separate steps and model each step with a deep neural network. We also use simple heuristics to guide the propagation of local textures from the boundary to the hole. We show that, by using such techniques, inpainting reduces to the problem of learning two image-feature translation functions in much smaller space and hence easier to train. We evaluate our method on several public datasets and show that we generate results of better visual quality than previous state-of-the-art methods.

Video prediction models based on convolutional networks, recurrent networks, and their combinations often result in blurry predictions. We identify an important contributing factor for imprecise predictions that has not been studied adequately in the literature: blind spots, i.e., lack of access to all relevant past information for accurately predicting the future. To address this issue, we introduce a fully context-aware architecture that captures the entire available past context for each pixel using Parallel Multi-Dimensional LSTM units and aggregates it using blending units. Our model outperforms a strong baseline network of 20 recurrent convolutional layers and yields state-of-the-art performance for next step prediction on three challenging real-world video datasets: Human 3.6M, Caltech Pedestrian, and UCF-101. Moreover, it does so with fewer parameters than several recently proposed models, and does not rely on deep convolutional networks, multi-scale architectures, separation of background and foreground modeling, motion flow learning, or adversarial training. These results highlight that full awareness of past context is of crucial importance for video prediction.

Current two-stage object detectors, which consists of a region proposal stage and a refinement stage, may produce unreliable results due to ill-localized proposed regions. To address this problem, we propose a context refinement algorithm that explores rich contextual information to better refine each proposed region. In particular, we first identify neighboring regions that may contain useful contexts and then perform refinement based on the extracted and unified contextual information. In practice, our method effectively improves the quality of the final detection results as well as region proposals. Empirical studies show that context refinement yields substantial and consistent improvements over different baseline detectors. Moreover, the proposed algorithm brings around 3% performance gain on PASCAL VOC benchmark and around 6% gain on MS COCO benchmark respectively.

While machine learning approaches to visual emotion recognition offer great promise, current methods consider training and testing models on small scale datasets covering limited visual emotion concepts. Our analysis identifies an important but long overlooked issue of existing visual emotion benchmarks in the form of dataset biases. We design a series of tests to show and measure how such dataset biases obstruct learning a generalizable emotion recognition model. Based on our analysis, we propose a webly supervised approach by leveraging a large quantity of stock image data. Our approach uses a simple yet effective curriculum guided training strategy for learning discriminative emotion features. We discover that the models learned using our large scale stock image dataset exhibit significantly better generalization ability than the existing datasets without the manual collection of even a single label. Moreover, visual representation learned using our approach holds a lot of promise across a variety of tasks on different image and video datasets.

Deep neural network compression has the potential to bring modern resource-hungry deep networks to resource-limited devices. However, in many of the most compelling deployment scenarios of compressed deep networks, the operational constraints matter: for example, a pedestrian detection network on a self-driving car may have to satisfy a latency constraint for safe operation. We propose the first principled treatment of deep network compression under operational constraints. We formulate the compression learning problem from the perspective of constrained Bayesian optimization, and introduce a cooling (annealing) strategy to guide the network compression towards the target constraints. Experiments on ImageNet demonstrate the value of modelling constraints directly in network compression.

We present COBLA---Constrained Optimization Based Low-rank Approximation---a systematic method of finding an optimal low-rank approximation of a trained convolutional neural network, subject to constraints in the number of multiply-accumulate (MAC) operations and the memory footprint. COBLA optimally allocates the constrained computation resource into each layer of the approximated network. The singular value decomposition of the network weight is computed, then a binary masking variable is introduced to denote whether a particular singular value and the corresponding singular vectors are used in low-rank approximation. With this formulation, the number of the MAC operations and the memory footprint are represented as linear constraints in terms of the binary masking variables. The resulted 0-1 integer programming problem is approximately solved by sequential quadratic programming. COBLA does not introduce any hyperparameter. We empirically demonstrate that COBLA outperforms prior art using the SqueezeNet and VGG-16 architecture on the ImageNet dataset.

Face recognition has witnessed great progresses in recent years, mainly attributed to the high-capacity model designed and the abundant labeled data collected. However, it becomes more and more prohibitive to scale up the current million-level identity annotations. In this work, we show that unlabeled face data can be as effective as the labeled ones. Here, we consider a setting closely mimicking the real-world scenario, where the unlabeled data are collected from unconstrained environment and their identities are exclusive from the labeled ones. Our main insight is that although the class information is not available, we can still faithfully approximate these semantic relationship by constructing a relational graph in a bottom-up manner. We propose Consensus-Driven Propagation (CDP) to tackle this challenging problem with two well-designed modules, the "committee" and the "mediator", which select positive face pairs robustly by carefully aggregating multi-view information. Extensive experiments validate the effectiveness of both modules to discard outliers and mine hard positives. With CDP, we achieve a compelling 78.18% on MegaFace identification challenge by using only 9% of the labels, comparing to 61.78% when no unlabeled data are used and 78.52% when all the labels are employed.

This paper addresses the challenging problem of estimating the general visual attention of people in images. Our proposed method is designed to work across multiple naturalistic social scenarios and provides a full picture of the subject’s attention and gaze. In contrast, earlier works on gaze and attention estimation have focused on constrained problems in more specific contexts. In particular, our model explicitly represents the gaze direction and handles out-of-frame gaze targets. We leverage three different datasets using a multi-task learning approach. We evaluate our method on widely used benchmarks for single-tasks such as gaze angle estimation and attention-within-an-image, as well as on the new challenging task of generalized visual attention prediction. In addition, we have created extended annotations for the MMDB and GazeFollow datasets which are used in our experiments, which we will publicly release.

Classical computation of optical flow involves generic priors (regularizers) that capture rudimentary statistics of images, but not long-range correlations or semantics. On the other hand, fully supervised methods learn the regularity in the annotated data, without explicit regularization and with the risk of overfitting. We seek to learn richer priors on the set of possible flows that are statistically compatible with an image. Once the prior is learned in a supervised fashion, one can easily learn the full map to infer optical flow directly from two or more images, without any need for (additional) supervision. We introduce a novel architecture, called Conditional Prior Network (CPN), and show how to train it to yield a conditional prior. When used in conjunction with a simple optical flow architecture, the CPN beats all variational method and all unsupervised learning-based ones. It performs comparably to fully supervised ones, that however are fine-tuned to a particular dataset. Our method, on the other hand, performs well even when transferred between datasets.

This paper presents an approach for grounding phrases in images which jointly learns multiple text-conditioned embeddings in a single end-to-end model. In order to differentiate text phrases into semantically distinct subspaces, we propose a concept weight branch that automatically assigns phrases to embeddings, whereas prior works predefine such assignments. Our proposed solution simplifies the representation requirements for individual embeddings and allows the underrepresented concepts to take advantage of the shared representations before feeding them into concept-specific layers. Comprehensive experiments verify the effectiveness of our approach across three phrase grounding datasets, Flickr30K Entities, ReferIt Game, and Visual Genome, where we obtain a (resp.) 4%, 3%, and 4% improvement in grounding performance over a strong region-phrase embedding baseline.

Existing works on semantic segmentation typically consider a small number of labels, ranging from tens to a few hundreds. With a large number of labels, training and evaluation of such task become extremely challenging due to correlation between labels and lack of datasets with complete annotations. We formulate semantic segmentation as a problem of image segmentation given a semantic concept, and propose a novel system which can potentially handle an unlimited number of concepts, including objects, parts, stuff, and attributes. We achieve this using a weakly and semi-supervised framework leveraging multiple datasets with different levels of supervision. We first train a deep neural network on a 6M stock image dataset with only image-level labels to learn visual-semantic embedding on 18K concepts. Then, we refine and extend the embedding network to predict an attention map, using a curated dataset with bounding box annotations on 750 concepts. Finally, we train an attention-driven class agnostic segmentation network using an 80-category fully annotated dataset. We perform extensive experiments to validate that the proposed system performs competitively to the state of the art on fully supervised concepts, and is capable of producing accurate segmentations for weakly learned and unseen concepts.

We consider the first-order scattering transform as a candidate for reducing the signal processed by a convolutional neural network (CNN). We study this transformation and show theoretical and empirical evidence that in the case of natural images and sufficiently small translation invariance, this transform preserves most of the signal information needed for classification while substantially reducing the spatial resolution and total signal size. We demonstrate that cascading a CNN with this representations permits to perform on par with Imagenet classification models commonly used in downstream tasks such as the Resnet-50. We subsequently apply our Imagenet trained hybrid model as a base model on a detection system, which typically has larger image inputs. On Pascal VOC and COCO detection tasks we find this leads to substantial improvements in inference speed and training memory consumption compared to models trained directly on the input image.

In this paper, we propose a new memory network structure for few-shot video classification by making the following contributions. First, we propose a compound memory network (CMN) structure under the key-value memory network paradigm, in which each key memory involves multiple constituent keys. These constituent keys work collaboratively for training, which enables the CMN to obtain an optimal video representation in a larger space. Second, we introduce a multi-saliency embedding algorithm which encodes a variable-length video sequence into a fixed-size matrix representation by discovering multiple saliencies of interest. For example, given a video of car auction, some people are interested in the car, while others are interested in the auction activities. Third, we design an abstract memory on top of the constituent keys. The abstract memory and constituent keys form a layered structure, which makes the CMN more efficient and capable of being scaled, while also retaining the representation capability of the multiple keys. We compare CMN with several state-of-the-art baselines on a new few-shot video classification dataset and show the effectiveness of our approach.

The world of human-object interactions is rich. While generally we sit on chairs and sofas, if need be we can even sit on TVs or top of shelves. In recent years, there has been progress in modeling actions and human-object interactions. However, most of these approaches require lots of data. It is not clear if the learned representations of actions are generalizable to new categories. In this paper, we explore the problem of zero-shot learning of human-object interactions. Given limited verb-noun interactions in training data, we want to learn a model than can work even on unseen combinations. To deal with this problem, In this paper, we propose a novel method using external knowledge graph and graph convolutional networks which learns how to compose classifiers for verb-noun pairs. We also provide benchmarks on several dataset for zero-shot learning including both image and video. We hope our method, dataset and baselines will facilitate future research in this direction.

With multiple crowd gatherings of millions of people every year in events ranging from pilgrimages to protests, concerts to marathons, and festivals to funerals; visual crowd analysis is emerging as a new frontier in computer vision. In particular, counting in highly dense crowds is a challenging problem with far-reaching applicability in crowd safety and management, as well as gauging political significance of protests and demonstrations. In this paper, we propose a novel approach that simultaneously solves the problems of counting, density map estimation and localization of people in a given dense crowd image. Our formulation is based on an important observation that the three problems are inherently related to each other making the loss function for optimizing a deep CNN decomposable. Since localization requires high-quality images and annotations, we introduce UCF-QNRF dataset that overcomes the shortcomings of previous datasets, and contains 1.25 million humans manually marked with dot annotations. Finally, we present evaluation measures and comparison with recent deep CNN networks, including those developed specifically for crowd counting. Our approach significantly outperforms state-of-the-art on the new dataset, which is the most challenging dataset with the largest number of crowd annotations in the most diverse set of scenes.