ICCV 2017 论文列表

Foreground map evaluation is crucial for gauging the progress of object segmentation algorithms, in particular in the filed of salient object detection where the purpose is to accurately detect and segment the most salient object in a scene. Several widely-used measures such as Area Under the Curve (AUC), Average Precision (AP) and the recently proposed Fbw have been utilized to evaluate the similarity between a non-binary saliency map (SM) and a ground-truth (GT) map. These measures are based on pixel-wise errors and often ignore the structural similarities. Behavioral vision studies, however, have shown that the human visual system is highly sensitive to structures in scenes. Here, we propose a novel, efficient, and easy to calculate measure known an structural similarity measure (Structure-measure) to evaluate non-binary foreground maps. Our new measure simultaneously evaluates region-aware and object-aware structural similarity between a SM and a GT map. We demonstrate superiority of our measure over existing ones using 5 meta-measures on 5 benchmark datasets.

We propose Stereo Direct Sparse Odometry (Stereo DSO) as a novel method for highly accurate real-time visual odometry estimation of large-scale environments from stereo cameras. It jointly optimizes for all the model parameters within the active window, including the intrinsic/extrinsic camera parameters of all keyframes and the depth values of all selected pixels. In particular, we propose a novel approach to integrate constraints from static stereo into the bundle adjustment pipeline of temporal multi-view stereo. Real-time optimization is realized by sampling pixels uniformly from image regions with sufficient intensity gradient. Fixed-baseline stereo resolves scale drift. It also reduces the sensitivities to large optical flow and to rolling shutter effect which are known shortcomings of direct image alignment methods. Quantitative evaluation demonstrates that the proposed Stereo DSO outperforms existing state-of-the-art visual odometry methods both in terms of tracking accuracy and robustness. Moreover, our method delivers a more precise metric 3D reconstruction than previous dense/semi-dense direct approaches while providing a higher reconstruction density than feature-based methods.

The intensive annotation cost and the rich but unlabeled data contained in videos motivate us to propose an unsupervised video-based person re-identification (re-ID) method. We start from two assumptions: 1) different video tracklets typically contain different persons, given that the tracklets are taken at distinct places or with long intervals; 2) within each tracklet, the frames are mostly of the same person. Based on these assumptions, this paper propose a stepwise metric promotion approach to estimate the identities of training tracklets, which iterates between cross-camera tracklet association and feature learning. Specifically, We use each training tracklet as a query, and perform retrieval in the cross camera training set. Our method is built on reciprocal nearest neighbor search and can eliminate the hard negative label matches, i.e., the cross-camera nearest neighbors of the false matches in the initial rank list. The tracklet that passes the reciprocal nearest neighbor check is considered to have the same ID with the query. Experimental results on the PRID 2011, ILIDS-VID, and MARS datasets show that the proposed method achieves very competitive re-ID accuracy compared with its supervised counterparts.

Synthesizing high-quality images from text descriptions is a challenging problem in computer vision and has many practical applications. Samples generated by existing text-to-image approaches can roughly reflect the meaning of the given descriptions, but they fail to contain necessary details and vivid object parts. In this paper, we propose Stacked Generative Adversarial Networks (StackGAN) to generate 256x256 photo-realistic images conditioned on text descriptions. We decompose the hard problem into more manageable sub-problems through a sketch-refinement process. The Stage-I GAN sketches the primitive shape and colors of the object based on the given text description, yielding Stage-I low-resolution images. The Stage-II GAN takes Stage-I results and text descriptions as inputs, and generates high-resolution images with photo-realistic details. It is able to rectify defects in Stage-I results and add compelling details with the refinement process. To improve the diversity of the synthesized images and stabilize the training of the conditional-GAN, we introduce a novel Conditioning Augmentation technique that encourages smoothness in the latent conditioning manifold. Extensive experiments and comparisons with state-of-the-arts on benchmark datasets demonstrate that the proposed method achieves significant improvements on generating photo-realistic images conditioned on text descriptions.

While strong progress has been made in image captioning recently, machine and human captions are still quite distinct. This is primarily due to the deficiencies in the generated word distribution, vocabulary size, and strong bias in the generators towards frequent captions. Furthermore, humans -- rightfully so -- generate multiple, diverse captions, due to the inherent ambiguity in the captioning task which is not explicitly considered in today's systems. To address these challenges, we change the training objective of the caption generator from reproducing ground-truth captions to generating a set of captions that is indistinguishable from human written captions. Instead of handcrafting such a learning target, we employ adversarial training in combination with an approximate Gumbel sampler to implicitly match the generated distribution to the human one. While our method achieves comparable performance to the state-of-the-art in terms of the correctness of the captions, we generate a set of diverse captions that are significantly less biased and better match the global uni-, bi- and tri-gram distributions of the human captions.

Region of Interest (ROI) crowd counting can be formulated as a regression problem of learning a mapping from an image or a video frame to a crowd density map. Recently, convolutional neural network (CNN) models have achieved promising results for crowd counting. However, even when dealing with video data, CNN-based methods still consider each video frame independently, ignoring the strong temporal correlation between neighboring frames. To exploit the otherwise very useful temporal information in video sequences, we propose a variant of a recent deep learning model called convolutional LSTM (ConvLSTM) for crowd counting. Unlike the previous CNN-based methods, our method fully captures both spatial and temporal dependencies. Furthermore, we extend the ConvLSTM model to a bidirectional ConvLSTM model which can access long-range information in both directions. Extensive experiments using four publicly available datasets demonstrate the reliability of our approach and the effectiveness of incorporating temporal information to boost the accuracy of crowd counting. In addition, we also conduct some transfer learning experiments to show that once our model is trained on one dataset, its learning experience can be transferred easily to a new dataset which consists of only very few video frames for model adaptation.

In this paper, we address the problem of spatio-temporal person retrieval from videos using a natural language query, in which we output a tube (i.e., a sequence of bounding boxes) which encloses the person described by the query. For this problem, we introduce a novel dataset consisting of videos containing people annotated with bounding boxes for each second and with five natural language descriptions. To retrieve the tube of the person described by a given natural language query, we design a model that combines methods for spatio-temporal human detection and multimodal retrieval. We conduct comprehensive experiments to compare a variety of tube and text representations and multimodal retrieval methods, and present a strong baseline in this task as well as demonstrate the efficacy of our tube representation and multimodal feature embedding technique. Finally, we demonstrate the versatility of our model by applying it to two other important tasks.

We aim for zero-shot localization and classification of human actions in video. Where traditional approaches rely on global attribute or object classification scores for their zero-shot knowledge transfer, our main contribution is a spatial-aware object embedding. To arrive at spatial awareness, we build our embedding on top of freely available actor and object detectors. Relevance of objects is determined in a word embedding space and further enforced with estimated spatial preferences. Besides local object awareness, we also embed global object awareness into our embedding to maximize actor and object interaction. Finally, we exploit the object positions and sizes in the spatial-aware embedding to demonstrate a new spatio-temporal action retrieval scenario with composite queries. Action localization and classification experiments on four contemporary action video datasets support our proposal. Apart from state-of-the-art results in the zero-shot localization and classification settings, our spatial-aware embedding is even competitive with recent supervised action localization alternatives.

Modeling instance-level context and object-object relationships is extremely challenging. It requires reasoning about bounding boxes of different locations, scales, aspect ratios etc.. Above all, instance-level spatial reasoning inherently requires modeling conditional distributions on previous detections. But our current object detection systems do not have any memory to remember what to condition on! The state-of-the-art object detectors still detect all object in parallel followed by non-maximal suppression (NMS). While memory has been used for tasks such as captioning and VQA, they use image-level memory cells without capturing the spatial layout. On the other hand, modeling object-object relationships requires spatial reasoning -- not only do we need a memory to store the spatial layout, but also a effective reasoning module to extract spatial patterns. This paper presents a conceptually simple yet powerful solution -- Spatial Memory Network (SMN), to model the instance-level context efficiently and effectively. Our spatial memory essentially assembles object instances back into a pseudo "image" representation that is easy to be fed into another ConvNet for object-object context reasoning. This leads to a new sequential reasoning architecture where image and memory are processed in parallel to obtain detections which update the memory again. We show our SMN architecture is effective as it provides 2.2% improvement over baseline Faster RCNN on the COCO dataset with VGG16.

Principal Component Analysis (PCA) is a widely popular dimensionality reduction technique for vector-valued inputs. In the past decade, a nonlinear generalization of PCA, called the Principal Geodesic Analysis (PGA) was developed to tackle data that lie on a smooth manifold. PGA suffers from the same problem as PCA in that, in both the methods, each Principal Component (PC) is a linear combination of the original variables. This makes it very difficult to interpret the PCs especially in high dimensions. This lead to the introduction of sparse PCA (SPCA) in the vector space input case. In this paper, we present a novel generalization of SPCA, called sparse exact PGA (SEPGA) that can cope with manifold-valued input data and respect the intrinsic geometry of the underlying manifold. Sparsity has the advantage of not only easy interpretability but also computational efficiency. We achieve this by formulating the PGA problem as a minimization of the projection error in conjunction with sparsity constraints enforced on the principal vectors post isomorphic mapping to Rm, where m is the dimension of the manifold on which the data reside. Further, for constant curvature smooth manifolds, we use analytic formulae for the projection error leading to an efficient solution to the SEPGA problem. We present extensive experimental results demonstrating the performance of SEPGA to achieve very good sparse principal components without sacrificing the accuracy of reconstruction. This makes the representation of manifold-valued data using SEPGA accurate and efficient.

This paper addresses the problem of building a spatio-temporal model of the world from a stream of time-stamped data. Unlike traditional models for simultaneous localization and mapping (SLAM) and structure-from-motion (SfM) which focus on recovering a single rigid 3D model, we tackle the problem of mapping scenes in which dynamic components appear, move and disappear independently of each other over time. We introduce a simple generative probabilistic model of 4D structure which specifies location, spatial and temporal extent of rigid surface patches by local Gaussian mixtures. We fit this model to a time-stamped stream of input data using expectation-maximization to estimate the model structure parameters (mapping) and the alignment of the input data to the model (localization). By explicitly representing the temporal extent and observability of surfaces in a scene, our method yields superior localization and reconstruction relative to baselines that assume a static 3D scene. We carry out experiments on both synthetic RGB-D data streams as well as challenging real-world datasets, tracking scene dynamics in a human workspace over the course of several weeks.

Non-maximum suppression is an integral part of the object detection pipeline. First, it sorts all detection boxes on the basis of their scores. The detection box M with the maximum score is selected and all other detection boxes with a significant overlap (using a pre-defined threshold) with M are suppressed. This process is recursively applied on the remaining boxes. As per the design of the algorithm, if an object lies within the predefined overlap threshold, it leads to a miss. To this end, we propose Soft-NMS, an algorithm which decays the detection scores of all other objects as a continuous function of their overlap with M. Hence, no object is eliminated in this process. Soft-NMS obtains consistent improvements for the coco-style mAP metric on standard datasets like PASCAL VOC 2007 (1.7% for both R-FCN and Faster-RCNN) and MS-COCO (1.3% for R-FCN and 1.1% for Faster-RCNN) by just changing the NMS algorithm without any additional hyper-parameters. Using Deformable-RFCN, Soft-NMS improves state-of-the-art in object detection from 39.8% to 40.9% with a single model. Further, the computational complexity of Soft-NMS is the same as traditional NMS and hence it can be efficiently implemented. Since Soft-NMS does not require any extra training and is simple to implement, it can be easily integrated into any object detection pipeline. Code for Soft-NMS is publicly available on GitHub

Weakly supervised object localization remains challenging, where only image labels instead of bounding boxes are available during training. Object proposal is an effective component in localization, but often computationally expensive and incapable of joint optimization with some of the remaining modules. In this paper, to the best of our knowledge, we for the first time integrate weakly supervised object proposal into convolutional neural networks (CNNs) in an end-to-end learning manner. We design a network component, Soft Proposal (SP), to be plugged into any standard convolutional architecture to introduce the nearly cost-free object proposal, orders of magnitude faster than state-of-the-art methods. In the SP-augmented CNNs, referred to as Soft Proposal Networks (SPNs), iteratively evolved object proposals are generated based on the deep feature maps then projected back, and further jointly optimized with network parameters, with image-level supervision only. Through the unified learning process, SPNs learn better object-centric filters, discover more discriminative visual evidence, and suppress background interference, significantly boosting both weakly supervised object localization and classification performance. We report the best results on popular benchmarks, including PASCAL VOC, MS COCO, and ImageNet.

To solve deep metric learning problems and produce feature embeddings, current methodologies will commonly use a triplet model to minimise the relative distance between samples from the same class and maximise the relative distance between samples from different classes. Though successful, the training convergence of this triplet model can be compromised by the fact that the vast majority of the training samples will produce gradients with magnitudes that are close to zero. This issue has motivated the development of methods that explore the global structure of the embedding and other methods that explore hard negative/positive mining. The effectiveness of such mining methods is often associated with intractable computational requirements. In this paper, we propose a novel deep metric learning method that combines the triplet model and the global structure of the embedding space. We rely on a smart mining procedure that produces effective training samples for a low computational cost. In addition, we propose an adaptive controller that automatically adjusts the smart mining hyper-parameters and speeds up the convergence of the training process. We show empirically that our proposed method allows for fast and more accurate training of triplet ConvNets than other competing mining methods. Additionally, we show that our method achieves new state-of-the-art embedding results for CUB-200-2011 and Cars196 datasets.

We propose a novel measure of visual similarity for image retrieval that incorporates both structural and aesthetic (style) constraints. Our algorithm accepts a query as sketched shape, and a set of one or more contextual images specifying the desired visual aesthetic. A triplet network is used to learn a feature embedding capable of measuring style similarity independent of structure, delivering significant gains over previous networks for style discrimination. We incorporate this model within a hierarchical triplet network to unify and learn a joint space from two discriminatively trained streams for style and structure. We demonstrate that this space enables, for the first time, style-constrained sketch search over a diverse domain of digital artwork comprising graphics, paintings and drawings. We also briefly explore alternative query modalities.

We address the problem of recognizing situations in images. Given an image, the task is to predict the most salient verb (action), and fill its semantic roles such as who is performing the action, what is the source and target of the action, etc. Different verbs have different roles (e.g. attacking has weapon), and each role can take on many possible values (nouns). We propose a model based on Graph Neural Networks that allows us to efficiently capture joint dependencies between roles using neural networks defined on a graph. Experiments with different graph connectivities show that our approach that propagates information between roles significantly outperforms existing work, as well as multiple baselines. We obtain roughly 3-5% improvement over previous work in predicting the full situation. We also provide a thorough qualitative analysis of our model and influence of different roles in the verbs.

We present a novel single-shot text detector that directly outputs word-level bounding boxes in a natural image. We propose an attention mechanism which roughly identifies text regions via an automatically learned attentional map. This substantially suppresses background interference in the convolutional features, which is the key to producing accurate inference of words, particularly at extremely small sizes. This results in a single model that essentially works in a coarse-to-fine manner. It departs from recent FCN-based text detectors which cascade multiple FCN models to achieve an accurate prediction. Furthermore, we develop a hierarchical inception module which efficiently aggregates multi-scale inception features. This enhances local details, and also encodes strong context information, allowing the detector to work reliably on multi-scale and multi-orientation text with single-scale images. Our text detector achieves an F-measure of 77% on the ICDAR 2015 benchmark, advancing the state-of-the-art results.

In this paper, we propose a novel single image action recognition algorithm based on the idea of semantic part actions. Unlike existing part-based methods, we argue that there exists a mid-level semantic, the semantic part action; and human action is a combination of semantic part actions and context cues. In detail, we divide human body into seven parts: head, torso, arms, hands and lower body. For each of them, we define a few semantic part actions (e.g.head: laughing). Finally, we exploit these part actions to infer the entire body action (e.g. applauding). To make the proposed idea practical, we propose a deep network-based framework which consists of two subnetworks, one for part localization and the other for action prediction. The action prediction network jointly learns part-level and body-level action semantics and combines them for the final decision. Extensive experiments demonstrate our proposal on semantic part actions as elements for entire body action. Our method reaches mAP of 93.9% and 91.2% on PASCAL VOC 2012 and Stanford-40, which outperforms the state-of-the-art by 2.3% and 8.6%.

Interestingly, shape of the high-altitude clouds serves as a beacon for weather forecasting, so its detection is of vital importance. Besides these clouds often cause hindrance in an endeavor of satellites to inspect our world. Even thin clouds produce the undesired superposition of visual information, whose decomposition into the clear background and cloudy layer using a single satellite image is a highly ill-posed problem. In this work, we derive sophisticated image priors by thoroughly analyzing the properties of high-altitude clouds and geological images; and formulate a non-convex optimization scheme, which simultaneously detects and removes the clouds within a few seconds. Experimental results on real world RGB images demonstrate that the proposed method outperforms the other competitive methods by retaining the comprehensive background details and producing the precise shape of the cloudy layer.