ECCV 2018 论文列表

Generating stylized captions for an image is an emerging topic in image captioning. Given an image as input, it requires the system to generate a caption that has a specific style (e.g., humorous, romantic, positive, and negative) while describing the image content semantically accurately. In this paper, we propose a novel stylized image captioning model that effectively takes both requirements into consideration. To this end, we first devise a new variant of LSTM, named style-factual LSTM, as the building block of our model. It uses two groups of matrices to capture the factual and stylized knowledge, respectively, and automatically learns the word-level weights of the two groups based on previous context. In addition, when we train the model to capture stylized elements, we propose an adaptive learning approach based on a reference factual model, it provides factual knowledge to the model as the model learns from stylized caption labels, and can adaptively compute how much information to supply at each time step. We evaluate our model on two stylized image captioning datasets, which contain humorous/romantic captions and positive/negative captions, respectively. Experiments shows that our proposed model outperforms the state-of-the-art approaches, without using extra ground truth supervision.

Recognizing visual relationships

Visual keyword spotting (KWS) is the problem of estimating whether a text query occurs in a given recording using only video information. This paper focuses on visual KWS for words unseen during training, a real-world, practical setting which so far has received no attention by the community. To this end, we devise an end-to-end architecture comprising (a) a state-of-the-art visual feature extractor based on spatiotemporal Residual Networks, (b) a grapheme-to-phoneme model based on sequence-to-sequence neural networks, and (c) a stack of recurrent neural networks which learn how to correlate visual features with the keyword representation. Different to prior works on KWS, which try to learn word representations merely from sequences of graphemes (i.e. letters), we propose the use of a grapheme-to-phoneme encoder-decoder model which learns how to map words to their pronunciation. We demonstrate that our system obtains very promising visual-only KWS results on the challenging LRS2 database, for keywords unseen during training. We also show that our system outperforms a baseline which addresses KWS via automatic speech recognition (ASR), while it drastically improves over other recently proposed ASR-free KWS methods.

We introduce and tackle the problem of zero-shot object detection (ZSD), which aims to detect object classes which are not observed during training. We work with a challenging set of object classes, not restricting ourselves to similar and/or fine-grained categories as in prior works on zero-shot classification. We present a principled approach by first adapting visual-semantic embeddings for ZSD. We then discuss the problems associated with selecting a background class and motivate two background-aware approaches for learning robust detectors. One of these models uses a fixed background class and the other is based on iterative latent assignments. We also outline the challenge associated with using a limited number of training classes and propose a solution based on dense sampling of the semantic label space using auxiliary data with a large number of categories. We propose novel splits of two standard detection datasets – MSCOCO and VisualGenome, and present extensive empirical results in both the traditional and generalized zero-shot settings to highlight the benefits of the proposed methods. We provide useful insights into the algorithm and conclude by posing some open questions to encourage further research.

Domain adaptation is an important tool to transfer knowledge about a task (e.g. classification) learned in a source domain to a second, or target domain. Current approaches assume that task-relevant target-domain data is available during training. We demonstrate how to perform domain adaptation when no such task-relevant target-domain data is available. To tackle this issue, we propose zero-shot deep domain adaptation (ZDDA), which uses privileged information from task-irrelevant dual-domain pairs. ZDDA learns a source-domain representation which is not only tailored for the task of interest but also close to the target-domain representation. Therefore, the source-domain task of interest solution (e.g. a classifier for classification tasks) which is jointly trained with the source-domain representation can be applicable to both the source and target representations. Using the MNIST, Fashion-MNIST, NIST, EMNIST, and SUN RGB-D datasets, we show that ZDDA can perform domain adaptation in classification tasks without access to task-relevant target-domain training data. We also extend ZDDA to perform sensor fusion in the SUN RGB-D scene classification task by simulating task-relevant target-domain representations with task-relevant source-domain data. To the best of our knowledge, ZDDA is the first domain adaptation and sensor fusion method which requires no task-relevant target-domain data. The underlying principle is not particular to computer vision data, but should be extensible to other domains.

Object detection is one of the major problems in computer vision, and has been extensively studied. Most of the existing detection works rely on labor-intensive supervision, such as ground truth bounding boxes of objects or at least image-level annotations. On the contrary, we propose an object detection method that does not require any form of human annotation on target tasks, by exploiting freely available web images. In order to facilitate effective knowledge transfer from web images, we introduce a multi-instance multi-label domain adaption learning framework with two key innovations. First of all, we propose an instance-level adversarial domain adaptation network with attention on foreground objects to transfer the object appearances from web domain to target domain. Second, to preserve the class-specific semantic structure of transferred object features, we propose a simultaneous transfer mechanism to transfer the supervision across domains through pseudo strong label generation. With our end-to-end framework that simultaneously learns a weakly supervised detector and transfers knowledge across domains, we achieved significant improvements over baseline methods on the benchmark datasets.

Learning long-term spatial-temporal features are critical for many video analysis tasks. However, existing video segmentation methods predominantly rely on static image segmentation techniques, and methods capturing temporal dependency for segmentation have to depend on pretrained optical flow models, leading to suboptimal solutions for the problem. End-to-end sequential learning to explore spatial-temporal features for video segmentation is largely limited by the scale of available video segmentation datasets, i.e., even the largest video segmentation dataset only contains 90 short video clips. To solve this problem, we build a new large-scale video object segmentation dataset called YouTube Video Object Segmentation dataset (YouTube-VOS). Our dataset contains 3,252 YouTube video clips and 78 categories including common objects and human activities. This is by far the largest video object segmentation dataset to our knowledge and we have released it at https://youtube-vos.org. Based on this dataset, we propose a novel sequence-to-sequence network to fully exploit long-term spatial-temporal information in videos for segmentation. We demonstrate that our method is able to achieve the best results on our YouTube-VOS test set and comparable results on DAVIS 2016 compared to the current state-of-the-art methods. Experiments show that the large scale dataset is indeed a key factor to the success of our model.

The objective of this paper is a neural network model that controls the pose and expression of a given face, using another face or modality (e.g. audio). This model can then be used for lightweight, sophisticated video and image editing. We make the following three contributions. First, we introduce a network, X2Face, that can control a source face (specified by one or more frames) using another face in a driving frame to produce a generated frame with the identity of the source frame but the pose and expression of the face in the driving frame. Second, we propose a method for training the network fully self-supervised using a large collection of video data. Third, we show that the generation process can be driven by other modalities, such as audio or pose codes, without any further training of the network. The generation results for driving a face with another face are compared to state-of-the-art self-supervised/supervised methods. We show that our approach is more robust than other methods, as it makes fewer assumptions about the input data. We also show examples of using our framework for video face editing.

In current Xray CT scanners, tomographic reconstruction relies only on directly transmitted photons. The models used for reconstruction have regarded photons scattered by the body as noise or disturbance to be disposed of, either by acquisition hardware (an anti-scatter grid) or by the reconstruction software. This increases the radiation dose delivered to the patient. Treating these scattered photons as a source of information, we solve an inverse problem based on a 3D radiative transfer model that includes both elastic (Rayleigh) and inelastic (Compton) scattering. We further present ways to make the solution numerically efficient. The resulting tomographic reconstruction is more accurate than traditional CT, while enabling significant dose reduction and chemical decomposition. Demonstrations include both simulations based on a standard medical phantom and a real scattering tomography experiment.

Most machine learning methods are known to capture and exploit biases of the training data. While some biases are beneficial for learning, others are harmful. Specifically, image captioning models tend to exaggerate biases present in training data (e.g., if a word is present in 60% of training sentences, it might be predicted in 70% of sentences at test time). This can lead to incorrect captions in domains where unbiased captions are desired, or required, due to over-reliance on the learned prior and image context. In this work we investigate generation of gender-specific caption words (e.g. man, woman) based on the person's appearance or the image context. We introduce a new Equalizer model that encourages equal gender probability when gender evidence is occluded in a scene and confident predictions when gender evidence is present. The resulting model is forced to look at a person rather than use contextual cues to make a gender-specific prediction. The losses that comprise our model, the Appearance Confusion Loss and the Confident Loss, are general, and can be added to any description model in order to mitigate impacts of unwanted bias in a description dataset. Our proposed model has lower error than prior work when describing images with people and mentioning their gender and more closely matches the ground truth ratio of sentences including women to sentences including men. Finally, we show that our model more often looks at people when predicting their gender.

Test datasets should contain many different challenging aspects so that the robustness and real-world applicability of algorithms can be assessed. In this work, we present a new test dataset for semantic and instance segmentation for the automotive domain. We have conducted a thorough risk analysis to identify situations and aspects that can reduce the output performance for these tasks. Based on this analysis we have designed our new dataset. Meta-information is supplied to mark which individual visual hazards are present in each test case. Furthermore, a new benchmark evaluation method is presented that uses the meta-information to calculate the robustness of a given algorithm with respect to the individual hazards. We show how this new approach allows for a more expressive characterization of algorithm robustness by comparing three baseline algorithms.

Object counting is an important task in computer vision due to its growing demand in applications such as surveillance, traffic monitoring, and counting everyday objects. State-of-the-art methods use regression-based optimization where they explicitly learn to count the objects of interest. These often perform better than detection-based methods that need to learn the more difficult task of predicting the location, size, and shape of each object. However, we propose a detection-based method that does not need to estimate the size and shape of the objects and that outperforms regression-based methods. Our contributions are three-fold: (1) we propose a novel loss function that encourages the network to output a single blob per object instance using point-level annotations only; (2) we design two methods for splitting large predicted blobs between object instances; and (3) we show that our method achieves new state-of-the-art results on several challenging datasets including the Pascal VOC and the Penguins dataset. Our method even outperforms those that use stronger supervision such as depth features, multi-point annotations, and bounding-box labels.

Simultaneously and accurately forecasting the behavior of many interacting agents is imperative for computer vision applications to be widely deployed (e.g., autonomous vehicles, security, surveillance, sports). In this paper, we present a technique using conditional variational autoencoder which learns a model that "personalizes'' prediction to individual agent behavior within a group representation. Given the volume of data available and its adversarial nature, we focus on the sport of basketball and show that our approach efficiently predicts context-specific agent motions. We find that our model generates results that are three times as accurate as previous state of the art approaches (5.74 ft vs. 17.95 ft).

Despite tremendous progress achieved in temporal action localization, state-of-the-art methods still struggle to train accurate models when annotated data is scarce. In this paper, we introduce a novel active learning framework for temporal localization that aims to mitigate this data dependency issue. We equip our framework with active selection functions that can reuse knowledge from previously annotated datasets. We study the performance of two state-of-the-art active selection functions as well as two widely used active learning baselines. To validate the effectiveness of each one of these selection functions, we conduct simulated experiments on ActivityNet. We find that using previously acquired knowledge as a bootstrapping source is crucial for active learners aiming to localize actions. When equipped with the right selection function, our proposed framework exhibits significantly better performance than standard active learning strategies, such as uncertainty sampling. Finally, we employ our framework to augment the newly compiled Kinetics action dataset with ground-truth temporal annotations. As a result, we collect Kinetics-Localization, a novel large-scale dataset for temporal action localization, which contains more than 15K YouTube videos.

Video summarization is a challenging under-constrained problem because the underlying summary of a single video strongly depends on users' subjective understandings. Data-driven approaches, such as deep neural networks, can deal with the ambiguity inherent in this task to some extent, but it is extremely expensive to acquire the temporal annotations of a large-scale video dataset. To leverage the plentiful web-crawled videos to improve the performance of video summarization, we present a generative modelling framework to learn the latent semantic video representations to bridge the benchmark data and web data. Specifically, our framework couples two important components: a variational autoencoder for learning the latent semantics from web videos, and an encoder-attention-decoder for saliency estimation of raw video and summary generation. A loss term to learn the semantic matching between the generated summaries and web videos is presented, and the overall framework is further formulated into a unified conditional variational encoder-decoder, called variational encoder-summarizer-decoder (VESD). Experiments conducted on the challenging datasets CoSum and TVSum demonstrate the superior performance of the proposed VESD to existing state-of-the-art methods. The source code of this work can be found at https://github.com/cssjcai/vesd.

Compared with depth-based 3D hand pose estimation, it is more challenging to infer 3D hand pose from monocular RGB images, due to substantial depth ambiguity and the difficulty of obtaining fully-annotated training data. Different from existing learning-based monocular RGB-input approaches that require accurate 3D annotations for training, we propose to leverage the depth images that can be easily obtained from commodity RGB-D cameras during training, while during testing we take only RGB inputs for 3D joint predictions. In this way, we alleviate the burden of the costly 3D annotations in real-world dataset. Particularly, we propose a weakly-supervised method, adaptating from fully-annotated synthetic dataset to weakly-labeled real-world dataset with the aid of a depth regularizer, which generates depth maps from predicted 3D pose and serves as weak supervision for 3D pose regression. Extensive experiments on benchmark datasets validate the effectiveness of the proposed depth regularizer in both weakly-supervised and fully-supervised settings.

We present a weakly supervised model that jointly performs both semantic- and instance-segmentation -- a particularly relevant problem given the substantial cost of obtaining pixel-perfect annotation for these tasks. In contrast to many popular instance segmentation approaches based on object detectors, our method does not predict any overlapping instances. Moreover, we are able to segment both ``thing'' and ``stuff'' classes, and thus explain all the pixels in the image. ``Thing'' classes are weakly-supervised with bounding boxes, and ``stuff'' with image-level tags. We obtain state-of-the-art results on Pascal VOC, for both full and weak supervision (which achieves about 95% of fully-supervised performance). Furthermore, we present the first weakly-supervised results on Cityscapes for both semantic- and instance-segmentation. Finally, we use our weakly supervised framework to analyse the relationship between annotation quality and predictive performance, which is of interest to dataset creators.

The Convolutional Neural Network (CNN) based region proposal generation method (i.e. region proposal network), trained using bounding box annotations, is an essential component in modern fully supervised object detectors. However, Weakly Supervised Object Detection (WSOD) has not benefited from CNN-based proposal generation due to the absence of bounding box annotations, and is relying on standard proposal generation methods such as selective search. In this paper, we propose a weakly supervised region proposal network which is trained using only image-level annotations. The weakly supervised region proposal network consists of two stages. The first stage evaluates the objectness scores of sliding window boxes by exploiting the low-level information in CNN and the second stage refines the proposals from the first stage using a region-based CNN classifier. Our proposed region proposal network is suitable for WSOD, can be plugged into a WSOD network easily, and can share its convolutional computations with the WSOD network. Experiments on the PASCAL VOC and ImageNet detection datasets show that our method achieves the state-of-the-art performance for WSOD with performance gain of about 3% on average.

In many domains of computer vision, generative adversarial networks (GANs) have achieved great success, among which the family of Wasserstein GANs (WGANs) is considered to be state-of-the-art due to the theoretical contributions and competitive qualitative performance. However, it is very challenging to approximate the k-Lipschitz constraint required by the Wasserstein-1 metric (W-met). In this paper, we propose a novel Wasserstein divergence (W-div), which is a relaxed version of W-met and does not require the k-Lipschitz constraint. As a concrete application, we introduce a Wasserstein divergence objective for GANs (WGAN-div), which can faithfully approximate W-div through optimization. Under various settings, including progressive growing training, we demonstrate the stability of the proposed WGAN-div owing to its theoretical and practical advantages over WGANs. Also, we study the quantitative and visual performance of WGAN-div on standard image synthesis benchmarks, showing the superior performance of WGAN-div compared to the state-of-the-art methods.