ECCV 2018 论文列表

This work proposes an algorithm, called NetAdapt, that automatically adapts a pre-trained deep neural network to a mobile platform given a resource budget. While many existing algorithms simplify networks based on the number of MACs or weights, optimizing those indirect metrics may not necessarily reduce the direct metrics, such as latency and energy consumption. To solve this problem, NetAdapt incorporates direct metrics into its adaptation algorithm. These direct metrics are evaluated using empirical measurements, so that detailed knowledge of the platform and toolchain is not required. NetAdapt automatically and progressively simplifies a pre-trained network until the resource budget is met while maximizing the accuracy. Experiment results show that NetAdapt achieves better accuracy versus latency trade-offs on both mobile CPU and mobile GPU, compared with the state-of-the-art automated network simplification algorithms. For image classification on the ImageNet dataset, NetAdapt achieves up to a 1.7x speedup in measured inference latency with equal or higher accuracy on MobileNets (V1&V2).

The automatic evaluation of image descriptions is an intricate task, and it is highly important in the development and fine-grained analysis of captioning systems. Existing metrics to automatically evaluate image captioning systems fail to achieve a satisfactory level of correlation with human judgements at the sentence-level. Moreover, these metrics, unlike humans, tend to focus on specific aspects of quality, such as the n-gram overlap or the semantic meaning. In this paper, we present the first learning-based metric to evaluate image captions. Our proposed framework enables us to incorporate both lexical and semantic information into a single learned metric. This results in an evaluator that takes into account various linguistic features to assess the caption quality. The experiments we performed to assess the proposed metric, show improvements upon the state of the art in terms of correlation with human judgements and demonstrate its superior robustness to distractions.

Several methods were recently proposed for the task of translating images between domains without prior knowledge in the form of correspondences. The existing methods apply adversarial learning to ensure that the distribution of the mapped source domain is indistinguishable from the target domain, which suffers from known stability issues. In addition, most methods rely heavily on ``cycle'' relationships between the domains, which enforce a one-to-one mapping. In this work, we introduce an alternative method: Non-Adversarial Mapping (NAM), which separates the task of target domain generative modeling from the cross-domain mapping task. NAM relies on a pre-trained generative model of the target domain, and aligns each source image with an image synthesized from the target domain, while jointly optimizing the domain mapping function. It has several key advantages: higher quality and resolution image translations, simpler and more stable training and reusable target models. Extensive experiments are presented validating the advantages of our method.

This paper presents a novel Mutual Learning to Adapt model (MuLA) for joint human parsing and pose estimation. It effectively exploits mutual benefits from both tasks and simultaneously boosts their performance. Different from existing post-processing or multi-task learning based methods, MuLA predicts dynamic task-specific model parameters via recurrently leveraging guidance information from its parallel tasks. Thus MuLA can fast adapt parsing and pose models to provide more powerful representations by incorporating information from their counterparts, giving more robust and accurate results. MuLA is implemented with convolutional neural networks and end-to-end trainable. Comprehensive experiments on benchmarks LIP and extended PASCAL-Person-Part demonstrate the effectiveness of the proposed MuLA model with superior performance to well established baselines.

We study an open problem of artwork identification and propose a new dataset dubbed Open Museum Identification Challenge (Open MIC). It contains photos of exhibits captured in 10 distinct exhibition spaces of several museums which showcase paintings, timepieces, sculptures, glassware, relics, science exhibits, natural history pieces, ceramics, pottery, tools and indigenous crafts. The goal of Open MIC is to stimulate research in domain adaptation, egocentric recognition and few-shot learning by providing a testbed complementary to the famous Office dataset which reaches ~90% accuracy. To form our dataset, we captured a number of images per art piece with a mobile phone and wearable cameras to form the source and target data splits, respectively. To achieve robust baselines, we build on a recent approach that aligns per-class scatter matrices of the source and target CNN streams. Moreover, we exploit the positive definite nature of such representations by using end-to-end Bregman divergences and the Riemannian metric. We present baselines such as training/evaluation per exhibition and training/evaluation on the combined set covering 866 exhibit identities. As each exhibition poses distinct challenges e.g., quality of lighting, motion blur, occlusions, clutter, viewpoint and scale variations, rotations, glares, transparency, non-planarity, clipping, we break down results w.r.t. these factors.

We present multiresolution tree-structured networks to process point clouds for 3D shape understanding and generation tasks. Our network represents a 3D shape as set of locality-preserving 1D ordered list of points at multiple resolutions. This allows efficient feed-forward processing through 1D convolutions, coarse-to-fine analysis through a multi-grid architecture, and it leads to faster convergence and small memory footprint during training. The proposed tree-structured encoders can be used to classify shapes and outperform existing point-based architectures on shape classification benchmarks, while tree-structured decoders can be used for generating point clouds directly and they outperform existing approaches for image-to-shape inference tasks learned using the ShapeNet dataset. Our model also allows unsupervised learning of point-cloud based shapes by using a variational autoencoder, leading to higher-quality generated shapes.

We develop using person gaze direction for scene understanding. In particular, we use intersecting gazes to learn 3D locations that people tend to look at, which is analogous to having multiple camera views. The 3D locations that we discover need not be visible to the camera. Conversely, knowing 3D locations of scene elements that draw visual attention, such as other people in the scene, can help infer gaze direction. We provide a Bayesian generative model for the temporal scene that captures the joint probability of camera parameters, locations of people, their gaze, what they are looking at, and locations of visual attention. Both the number of people in the scene and the number of extra objects that draw attention are unknown and need to be inferred. To execute this joint inference we use a probabilistic data association approach that enables principled comparison of model hypotheses. We use MCMC for inference over the discrete correspondence variables, and approximate the marginalization over continuous parameters using the Metropolis-Laplace approximation, using Hamiltonian (Hybrid) Monte Carlo for maximization. As existing data sets do not provide the 3D locations of what people are looking at, we contribute a small data set that does. On this data set, we infer what people are looking at with 59% precision compared with 13% for a baseline approach, and where those objects are within about 0.58m.

We tackle here the problem of multimodal image non-rigid registration, which is of prime importance in remote sensing and medical imaging. The difficulties encountered by classical registration approaches include feature design and slow optimization by gradient descent. By analyzing these methods, we note the significance of the notion of scale. We design easy-to-train, fully-convolutional neural networks able to learn scale-specific features. Once chained appropriately, they perform global registration in linear time, getting rid of gradient descent schemes by predicting directly the deformation. We show their performance in terms of quality and speed through various tasks of remote sensing multimodal image alignment. In particular, we are able to register correctly cadastral maps of buildings as well as road polylines onto RGB images, and outperform current keypoint matching methods.

Unsupervised image-to-image translation is an important and challenging problem in computer vision. Given an image in the source domain, the goal is to learn the conditional distribution of corresponding images in the target domain, without seeing any pairs of corresponding images. While this conditional distribution is inherently multimodal, existing approaches make an overly simplified assumption, modeling it as a deterministic one-to-one mapping. As a result, they fail to generate diverse outputs from a given source domain image. To address this limitation, we propose a Multimodal Unsupervised Image-to-image Translation (MUNIT) framework. We assume that the image representation can be decomposed into a content code that is domain-invariant, and a style code that captures domain-specific properties. To translate an image to another domain, we recombine its content code with a random style code sampled from the style space of the target domain. We analyze the proposed framework and establish several theoretical results. Extensive experiments with comparisons to the state-of-the-art approaches further demonstrates the advantage of the proposed framework. Moreover, our framework allows users to control the style of translation outputs by providing an example style image.

We propose a video story question-answering (QA) architecture, Multimodal Dual Attention Memory (MDAM). The key idea is to use a dual attention mechanism with late fusion. MDAM uses self-attention to learn the latent concepts in scene frames and captions. Given a question, MDAM uses the second attention over these latent concepts. Multimodal fusion is performed after the dual attention processes (late fusion). Using this processing pipeline, MDAM learns to infer a high-level vision-language joint representation from an abstraction of the full video content. We evaluate MDAM on PororoQA and MovieQA datasets which have large-scale QA annotations on cartoon videos and movies, respectively. For both datasets, MDAM achieves new state-of-the-art results with significant margins compared to the runner-up models. We confirm the best performance of the dual attention mechanism combined with late fusion by ablation studies. We also perform qualitative analysis by visualizing the inference mechanisms of MDAM.

In this paper, we present MultiPoseNet, a novel bottom-up multi-person pose estimation architecture that combines a multi-task model with a novel assignment method. MultiPoseNet can jointly handle person detection, person segmentation and pose estimation problems. The novel assignment method is implemented by the Pose Residual Network (PRN) which receives keypoint and person detections, and produces accurate poses by assigning keypoints to person instances. On the COCO keypoints dataset, our pose estimation method outperforms all previous bottom-up methods both in accuracy (+4-point mAP over previous best result) and speed; it also performs on par with the best top-down methods while being at least 4x faster. Our method is the fastest real time system with ~23 frames/sec.

In this paper, we address the task of multi-view novel view synthesis, where we are interested in synthesizing a target image with an arbitrary camera pose from given source images. We propose an end-to-end trainable framework that learns to exploit multiple viewpoints to synthesize a novel view without any 3D supervision. Specifically, our model consists of a flow prediction module and a pixel generation module to directly leverage information presented in source views as well as hallucinate missing pixels from statistical priors. To merge the predictions produced by the two modules given multi-view source images, we introduce a self-learned confidence aggregation mechanism. We evaluate our model on images rendered from 3D object models as well as real and synthesized scenes. We demonstrate that our model is able to achieve state-of-the-art results as well as progressively improve its predictions when more source images are available.

Recent studies have shown that deep neural networks can significantly improve the quality of single-image super-resolution. Current researches tend to use deeper convolutional neural networks to enhance performance. However, blindly increasing the depth of the network cannot ameliorate the network effectively. Worse still, with the depth of the network increases, more problems occurred in the training process and more training tricks are needed. In this paper, we propose a novel multi-scale residual network (MSRN) to fully exploit the image features, which outperform most of the state-of-the-art methods. Based on the residual block, we introduce convolution kernels of different sizes to adaptively detect the image features in different scales. Meanwhile, we let these features interact with each other to get the most efficacious image information, we call this structure Multi-scale Residual Block (MSRB). Furthermore, the outputs of each MSRB are used as the hierarchical features for global feature fusion. Finally, all these features are sent to the reconstruction module for recovering the high-quality image.

In recent deep online and near-online multi-object tracking approaches, a difficulty has been to incorporate long-term appearance models to efficiently score object tracks under severe occlusion and multiple missing detections. In this paper, we propose a novel recurrent network model, the bilinear LSTM, in order to improve long-term appearance models via a recurrent network. Based on intuitions drawn from recursive least squares, bilinear LSTM stores building blocks of a linear predictor in its memory, which is then coupled with the input in a multiplicative manner, instead of the additive coupling in conventional LSTM approaches. Such coupling resembles an online learned classifier/regressor at each time step, which we have found to improve performances in using LSTM for appearance modeling. We also propose novel data augmentation approaches to efficiently train recurrent models that score object tracks on both appearance and motion. We train an LSTM that can score object tracks based on both appearance and motion and utilize it in a multiple hypothesis tracking framework. In experiments, we show that with our novel LSTM model, we achieved state-of-the-art performance on near-online multiple object tracking on the MOT 2016 and MOT 2017 benchmarks.

In generalized zero shot learning (GZSL), the set of classes are split into seen and unseen classes, where training relies on the semantic features of the seen and unseen classes and the visual representations of only the seen classes, while testing uses the visual representations of the seen and unseen classes. Current methods address GZSL by learning a transformation from the visual to the semantic space, exploring the assumption that the distribution of classes in the semantic and visual spaces is relatively similar. Such methods tend to transform unseen testing visual representations into one of the seen classes' semantic features instead of the semantic features of the correct unseen class, resulting in low accuracy GZSL classification. Recently, generative adversarial networks (GAN) have been explored to synthesize visual representations of the unseen classes from their semantic features - the synthesized representations of the seen and unseen classes are then used to train the GZSL classifier. This approach has been shown to boost GZSL classification accuracy, but there is one important missing constraint: there is no guarantee that synthetic visual representations can generate back their semantic feature in a multi-modal cycle-consistent manner. This missing constraint can result in synthetic visual representations that do not represent well their semantic features, which means that the use of this constraint can improve GAN-based approaches. In this paper, we propose the use of such constraint based on a new regularization for the GAN training that forces the generated visual features to reconstruct their original semantic features. Once our model is trained with this multi-modal cycle-consistent semantically, we can then synthesize more representative visual features for the seen and, more importantly, for the unseen classes. Our proposed approach shows the best GZSL classification results in the field in several publicly available datasets.

We develop a robust multi-scale structure-aware neural network for human pose estimation. This method improves the recent deep conv-deconv hourglass models with four key improvements: (1) multi-scale supervision to strengthen contextual feature learning in matching body keypoints by combining feature heatmaps across scales, (2) multi-scale regression network at the end to globally optimize the structural matching of the multi-scale features, (3) structure-aware loss used in the intermediate supervision and at the regression to improve the matching of keypoints and respective neighbors to infer a higher-order matching configurations, and (4) a keypoint masking training scheme that can effectively fine-tune our network to robustly localize occluded keypoints via adjacent matches. Our method can effectively improve state-of-the-art pose estimation methods that suffer from difficulties in scale varieties, occlusions, and complex multi-person scenarios. This multi-scale supervision tightly integrates with the regression network to effectively (i) localize keypoints using the ensemble of multi-scale features, and (ii) infer global pose configuration by maximizing structural consistencies across multiple keypoints and scales. The keypoint masking training enhances these advantages to focus learning on hard occlusion samples. Our method achieves the leading position in the MPII challenge leaderboard among the state-of-the-art methods.

Convolution is spatially-symmetric, i.e., the visual features are independent of its position in the image, which limits its ability to use spatial information. This paper addresses this issue by a recalibration process, which refers to the surrounding region of each neuron, computes an importance value and multiplies it to the original neural response. Our approach is named multi-scale spatially-asymmetric recalibration (MS-SAR), which, besides introducing spatial asymmetry into convolution, extracts visual cues from regions at multiple scales to allow richer information to be incorporated. MS-SAR is implemented in an efficient way, so that only small fractions of extra parameters and computations are required. We apply MS-SAR to several popular network architectures, in which all convolutional layers are recalibrated, and demonstrate superior performance in both CIFAR and ILSVRC2012 classification tasks.

Accurate semantic image segmentation requires the joint consideration of local appearance, semantic information, and global scene context. In today’s age of pre-trained deep networks and their powerful convolutional features, state-of-the-art semantic segmentation approaches differ mostly in how they choose to combine together these different kinds of information. In this work, we propose a novel scheme for aggregating features from different scales, which we refer to as Multi-Scale Context Intertwining (MSCI). In contrast to previous approaches, which typically propagate information between scales in a one-directional manner, we merge pairs of feature maps in a bidirectional and recurrent fashion, via connections between two LSTM chains. By training the parameters of the LSTM units on the segmentation task, the above approach learns how to extract powerful and effective features for pixel-level semantic segmentation, which are then combined hierarchically. Furthermore, rather than using fixed information propagation routes, we subdivide images into super-pixels, and use the spatial relationship between them in order to perform image-adapted context aggregation. Our extensive evaluation on public benchmarks indicates that all of the aforementioned components of our approach increase the effectiveness of information propagation throughout the network, and significantly improve its eventual segmentation accuracy.

In this paper, we aim to reduce the computational cost of spatio-temporal deep neural networks, making them run as fast as their 2D counterparts while preserving state-of-the-art accuracy on video recognition benchmarks. To this end, we present the novel Multi-Fiber architecture that slices a complex neural network into an ensemble of lightweight networks or fibers that run through the network. To facilitate information flow between fibers we further incorporate multiplexer modules and end up with an architecture that reduces the computational cost of 3D networks by an order of magnitude, while increasing recognition performance at the same time. Extensive experimental results show that our multi-fiber architecture significantly boosts the efficiency of existing convolution networks for both image and video recognition tasks, achieving state-of-the-art performance on UCF-101, HMDB-51 and Kinetics datasets. Our proposed model requires over 9× and 13× less computations than the I3D and R(2+1)D models, respectively, yet providing higher accuracy.