arXiv subjects

Yi Yang

Publications and source records attributed to Yi Yang.

At least 1,081 records · Page 60Linked to original sources

arXiv2017

Twitter100k: A Real-world Dataset for Weakly Supervised Cross-Media Retrieval

This paper contributes a new large-scale dataset for weakly supervised cross-media retrieval, named Twitter100k. Current datasets, such as Wikipedia, NUS Wide and Flickr30k, have two major limitations. First, these datasets are lacking in content diversity, i.e., only some pre-defined classes are covered. Second, texts in these datasets are written in well-organized language, leading to inconsistency with realistic applications. To overcome these drawbacks, the proposed Twitter100k dataset is characterized by two aspects: 1) it has 100,000 image-text pairs randomly crawled from Twitter and thus has no constraint in the image categories; 2) text in Twitter100k is written in informal language by the users. Since strongly supervised methods leverage the class labels that may be missing in practice, this paper focuses on weakly supervised learning for cross-media retrieval, in which only text-image pairs are exploited during training. We extensively benchmark the performance of four subspace learning methods and three variants of the Correspondence AutoEncoder, along with various text features on Wikipedia, Flickr30k and Twitter100k. Novel insights are provided. As a minor contribution, inspired by the characteristic of Twitter100k, we propose an OCR-based cross-media retrieval method. In experiment, we show that the proposed OCR-based method improves the baseline performance.

cs.CV↗

arXiv2017

A New Evaluation Protocol and Benchmarking Results for Extendable Cross-media Retrieval

This paper proposes a new evaluation protocol for cross-media retrieval which better fits the real-word applications. Both image-text and text-image retrieval modes are considered. Traditionally, class labels in the training and testing sets are identical. That is, it is usually assumed that the query falls into some pre-defined classes. However, in practice, the content of a query image/text may vary extensively, and the retrieval system does not necessarily know in advance the class label of a query. Considering the inconsistency between the real-world applications and laboratory assumptions, we think that the existing protocol that works under identical train/test classes can be modified and improved. This work is dedicated to addressing this problem by considering the protocol under an extendable scenario, \ie, the training and testing classes do not overlap. We provide extensive benchmarking results obtained by the existing protocol and the proposed new protocol on several commonly used datasets. We demonstrate a noticeable performance drop when the testing classes are unseen during training. Additionally, a trivial solution, \ie, directly using the predicted class label for cross-media retrieval, is tested. We show that the trivial solution is very competitive in traditional non-extendable retrieval, but becomes less so under the new settings. The train/test split, evaluation code, and benchmarking results are publicly available on our website.

cs.CV↗

arXiv2017

A Discriminatively Learned CNN Embedding for Person Re-identification

We revisit two popular convolutional neural networks (CNN) in person re-identification (re-ID), i.e, verification and classification models. The two models have their respective advantages and limitations due to different loss functions. In this paper, we shed light on how to combine the two models to learn more discriminative pedestrian descriptors. Specifically, we propose a new siamese network that simultaneously computes identification loss and verification loss. Given a pair of training images, the network predicts the identities of the two images and whether they belong to the same identity. Our network learns a discriminative embedding and a similarity measurement at the same time, thus making full usage of the annotations. Albeit simple, the learned embedding improves the state-of-the-art performance on two public person re-ID benchmarks. Further, we show our architecture can also be applied in image retrieval.

cs.CV↗

arXiv2017

The Sizes and Depletions of the Dust and Gas Cavities in the Transitional Disk J160421.7-213028

We report ALMA Cycle 2 observations of 230 GHz (1.3 mm) dust continuum emission, and $^{12}$CO, $^{13}$CO, and C$^{18}$O J = 2-1 line emission, from the Upper Scorpius transitional disk [PZ99] J160421.7-213028, with an angular resolution of ~0".25 (35 AU). Armed with these data and existing H-band scattered light observations, we measure the size and depth of the disk's central cavity, and the sharpness of its outer edge, in three components: sub-$μ$m-sized "small" dust traced by scattered light, millimeter-sized "big" dust traced by the millimeter continuum, and gas traced by line emission. Both dust populations feature a cavity of radius $\sim$70 AU that is depleted by factors of at least 1000 relative to the dust density just outside. The millimeter continuum data are well explained by a cavity with a sharp edge. Scattered light observations can be fitted with a cavity in small dust that has either a sharp edge at 60 AU, or an edge that transitions smoothly over an annular width of 10 AU near 60 AU. In gas, the data are consistent with a cavity that is smaller, about 15 AU in radius, and whose surface density at 15 AU is $10^{3\pm1}$ times smaller than the surface density at 70 AU; the gas density grades smoothly between these two radii. The CO isotopologue observations rule out a sharp drop in gas surface density at 30 AU or a double-drop model as found by previous modeling. Future observations are needed to assess the nature of these gas and dust cavities, e.g., whether they are opened by multiple as-yet-unseen planets or photoevaporation.

astro-ph.SR↗

arXiv2017

Pose Invariant Embedding for Deep Person Re-identification

Pedestrian misalignment, which mainly arises from detector errors and pose variations, is a critical problem for a robust person re-identification (re-ID) system. With bad alignment, the background noise will significantly compromise the feature learning and matching process. To address this problem, this paper introduces the pose invariant embedding (PIE) as a pedestrian descriptor. First, in order to align pedestrians to a standard pose, the PoseBox structure is introduced, which is generated through pose estimation followed by affine transformations. Second, to reduce the impact of pose estimation errors and information loss during PoseBox construction, we design a PoseBox fusion (PBF) CNN architecture that takes the original image, the PoseBox, and the pose estimation confidence as input. The proposed PIE descriptor is thus defined as the fully connected layer of the PBF network for the retrieval task. Experiments are conducted on the Market-1501, CUHK03, and VIPeR datasets. We show that PoseBox alone yields decent re-ID accuracy and that when integrated in the PBF network, the learned PIE descriptor produces competitive performance compared with the state-of-the-art approaches.

cs.CV↗

arXiv2017

Subaru/SCExAO First-Light Direct Imaging of a Young Debris Disk around HD 36546

We present $H$-band scattered light imaging of a bright debris disk around the A0 star HD 36546 obtained from the Subaru Coronagraphic Extreme Adaptive Optics (SCExAO) system with data recorded by the HiCIAO camera using the vector vortex coronagraph. SCExAO traces the disk from $r$ $\sim$ 0.3" to $r$ $\sim$ 1" (34--114 au). The disk is oriented in a near east-west direction (PA $\sim$ 75$^{o}$), is inclined by $i$ $\sim$ 70--75$^{o}$ and is strongly forward-scattering ($g$ $>$ 0.5). It is an extended disk rather than a sharp ring; a second, diffuse dust population extends from the disk's eastern side. While HD 36546 intrinsic properties are consistent with a wide age range ($t$ $\sim$ 1--250 $Myr$), its kinematics and analysis of coeval stars suggest a young age (3--10 $Myr$) and a possible connection to Taurus-Auriga's star formation history. SCExAO's planet-to-star contrast ratios are comparable to the first-light Gemini Planet Imager contrasts; for an age of 10 $Myr$, we rule out planets with masses comparable to HR 8799 b beyond a projected separation of 23 au. A massive icy planetesimal disk or an unseen superjovian planet at $r$ $>$ 20 au may explain the disk's visibility. The HD 36546 debris disk may be the youngest debris disk yet imaged, is the first newly-identified object from the now-operational SCExAO extreme AO system, is ideally suited for spectroscopic follow up with SCExAO/CHARIS in 2017, and may be a key probe of icy planet formation and planet-disk interactions.

astro-ph.EP↗

arXiv2016

Low-loss Plasmonic Dielectric Nanoresonators

Material losses in metals are a central bottleneck in plasmonics for many applications. Here we propose and theoretically demonstrate that metal losses can be successfully mitigated with dielectric particles on metallic films, giving rise to hybrid dielectric--metal resonances. In the far field, they yield strong and efficient scattering, beyond even the theoretical limits of all-metal and all-dielectric structures. In the near field, they offer high-Purcell-factor (${>}5000$), high-quantum-efficiency (${>}90\%$), and highly directional emission at visible and infrared wavelengths. Their quality factors can be readily tailored from plasmonic-like (${\sim}10$) to dielectric-like (${\sim}10^3$), with wide control over the individual resonant coupling to photon, plasmon, and dissipative channels. Compared with conventional plasmonic nanostructures, such resonances show robustness against detrimental nonlocal effects and provide higher field enhancement at extreme nanoscopic sizes and spacings. These hybrid resonances equip plasmonics with high efficiency, which has been the predominant goal since the field's inception.

physics.optics↗

arXiv2016

Review on High energy String Scattering Amplitudes and Symmetries of String Theory

We review high energy symmetries of string theory at both the fixed angle or Gross regime (GR) and the fixed momentum transfer or Regge regime (RR). We calculated in details high energy string scattering amplitudes at arbitrary mass levels for both regimes. We discovered infinite linear relations among fixed angle string amplitudes conjectured by Gross in 1988 from decoupling of high energy zero-norm states (ZNS), and infinite recurrence relations among Regge string amplitudes from Kummer function U and Appell function F_1. However, the linear relations we obtained in the GR corrected [27-32] the saddle point calculations of Gross, Gross and Mende and Gross and Manes [1-5]. Our results were consistent with the decoupling of high energy ZNS or unitarity of the theory while those of them were not. In addition, for the case of high energy closed string scatterings, our results [36] differ from theirs by an oscillating prefactor which was crucial to recover the KLT relation valid for all energies. In the GR/RR regime, all high energy string amplitudes can be solved by these linear/recurrence relations so that all GR/RR string amplitudes can be expressed in terms of one single GR/RR string amplitude. In addition, we found an interesting link between string amplitudes of the two regimes, and discovered that at each mass level the ratios among fixed angle amplitudes can be extracted from Regge string scattering amplitudes. This result enables us to argue that the known SL(5,C) dynamical symmetry of the Appell function F_1 is crucial to probe high energy spacetime symmetry of string theory.

hep-th↗

arXiv2016

Few-Shot Object Recognition from Machine-Labeled Web Images

With the tremendous advances of Convolutional Neural Networks (ConvNets) on object recognition, we can now obtain reliable enough machine-labeled annotations easily by predictions from off-the-shelf ConvNets. In this work, we present an abstraction memory based framework for few-shot learning, building upon machine-labeled image annotations. Our method takes some large-scale machine-annotated datasets (e.g., OpenImages) as an external memory bank. In the external memory bank, the information is stored in the memory slots with the form of key-value, where image feature is regarded as key and label embedding serves as value. When queried by the few-shot examples, our model selects visually similar data from the external memory bank, and writes the useful information obtained from related external data into another memory bank, i.e., abstraction memory. Long Short-Term Memory (LSTM) controllers and attention mechanisms are utilized to guarantee the data written to the abstraction memory is correlated to the query example. The abstraction memory concentrates information from the external memory bank, so that it makes the few-shot recognition effective. In the experiments, we firstly confirm that our model can learn to conduct few-shot object recognition on clean human-labeled data from ImageNet dataset. Then, we demonstrate that with our model, machine-labeled image annotations are very effective and abundant resources to perform object recognition on novel categories. Experimental results show that our proposed model with machine-labeled annotations achieves great performance, only with a gap of 1% between of the one with human-labeled annotations.

cs.CV↗

arXiv2016

Bidirectional Multirate Reconstruction for Temporal Modeling in Videos

Despite the recent success of neural networks in image feature learning, a major problem in the video domain is the lack of sufficient labeled data for learning to model temporal information. In this paper, we propose an unsupervised temporal modeling method that learns from untrimmed videos. The speed of motion varies constantly, e.g., a man may run quickly or slowly. We therefore train a Multirate Visual Recurrent Model (MVRM) by encoding frames of a clip with different intervals. This learning process makes the learned model more capable of dealing with motion speed variance. Given a clip sampled from a video, we use its past and future neighboring clips as the temporal context, and reconstruct the two temporal transitions, i.e., present$\rightarrow$past transition and present$\rightarrow$future transition, reflecting the temporal information in different views. The proposed method exploits the two transitions simultaneously by incorporating a bidirectional reconstruction which consists of a backward reconstruction and a forward reconstruction. We apply the proposed method to two challenging video tasks, i.e., complex event detection and video captioning, in which it achieves state-of-the-art performance. Notably, our method generates the best single feature for event detection with a relative improvement of 10.4% on the MEDTest-13 dataset and achieves the best performance in video captioning across all evaluation metrics on the YouTube2Text dataset.

cs.CV↗

arXiv2016

Near-Infrared Imaging Polarimetry of Inner Region of GG Tau A Disk

By performing non-masked polarization imaging with Subaru/HiCIAO, polarized scattered light from the inner region of the disk around the GG Tau A system was successfully detected in the $H$ band with a spatial resolution of approximately 0.07$\arcsec$, revealing the complicated inner disk structures around this young binary. This paper reports the observation of an arc-like structure to the north of GG Tau Ab and part of a circumstellar structure that is noticeable around GG Tau Aa extending to a distance of approximately 28 AU from the primary star. The speckle noise around GG Tau Ab constrains its disk radius to <13 AU. Based on the size of the circumbinary ring and the circumstellar disk around GG Tau Aa, the semi-major axis of the binary's orbit is likely to be 62 AU. A comparison of the present observations with previous ALMA and near-infrared (NIR) H$_2$ emission observations suggests that the north arc could be part of a large streamer flowing from the circumbinary ring to sustain the circumstellar disks. According to the previous studies, the circumstellar disk around GG Tau Aa has enough mass and can sustain itself for a duration sufficient for planet formation; thus, our study indicates that planets can form within close (separation $\lesssim$ 100 AU) young binary systems.

astro-ph.EP↗

arXiv2016

Interstellar-Medium Mapping in M82 Through Light Echoes Around Supernova 2014J

We present multiple-epoch measurements of the size and surface brightness of the light echoes from supernova (SN) 2014J in the nearby starburst galaxy M82. Hubble Space Telescope (HST) ACS/WFC images were taken ~277 and ~416 days after B-band maximum in the filters F475W, F606W, and F775W. Observations with HST WFC3/UVIS images at epochs ~216 and ~365 days (Crotts 2015) are included for a more complete analysis. The images reveal the temporal evolution of at least two major light-echo components. The first one exhibits a filled ring structure with position-angle-dependent intensity. This radially extended, diffuse echo indicates the presence of an inhomogeneous interstellar dust cloud ranging from ~100 pc to ~500 pc in the foreground of the SN. The second echo component appears as an unresolved luminous quarter-circle arc centered on the SN. The wavelength dependence of scattering measured in different dust components suggests that the dust producing the luminous arc favors smaller grain sizes, while that causing the diffuse light echo may have sizes similar to those of the Milky Way dust. Smaller grains can produce an optical depth consistent with that along the supernova-Earth line of sight measured by previous studies around maximum light. Therefore, it is possible that the dust slab, from which the luminous arc arises, is also responsible for most of the extinction towards SN 2014J. The optical depths determined from the Milky Way-like dust in the scattering matters are lower than that produced by the dust slab.

astro-ph.HE↗

arXiv2016

A Resolved Near-Infrared Image of The Inner Cavity in The GM Aur Transitional Disk

We present high-contrast H-band polarized intensity (PI) images of the transitional disk around the young solar-like star GM Aur. The near-infrared direct imaging of the disk was derived by polarimetric differential imaging using the Subaru 8.2-m Telescope and HiCIAO. An angular resolution and an inner working angle of 0."07 and r~0."05, respectively, were obtained. We clearly resolved a large inner cavity, with a measured radius of 18+/-2 au, which is smaller than that of a submillimeter interferometric image (28 au). This discrepancy in the cavity radii at near-infrared and submillimeter wavelengths may be caused by a 3-4M_Jup planet about 20 au away from the star, near the edge of the cavity. The presence of a near-infrared inner is a strong constraint on hypotheses for inner cavity formation in a transitional disk. A dust filtration mechanism has been proposed to explain the large cavity in the submillimeter image, but our results suggest that this mechanism must be combined with an additional process. We found that the PI slope of the outer disk is significantly different from the intensity slope obtained from HST/NICMOS, and this difference may indicate the grain growth process in the disk.

astro-ph.SR↗

arXiv2016

Optimizing Memory Efficiency for Deep Convolutional Neural Networks on GPUs

Leveraging large data sets, deep Convolutional Neural Networks (CNNs) achieve state-of-the-art recognition accuracy. Due to the substantial compute and memory operations, however, they require significant execution time. The massive parallel computing capability of GPUs make them as one of the ideal platforms to accelerate CNNs and a number of GPU-based CNN libraries have been developed. While existing works mainly focus on the computational efficiency of CNNs, the memory efficiency of CNNs have been largely overlooked. Yet CNNs have intricate data structures and their memory behavior can have significant impact on the performance. In this work, we study the memory efficiency of various CNN layers and reveal the performance implication from both data layouts and memory access patterns. Experiments show the universal effect of our proposed optimizations on both single layers and various networks, with up to 27.9x for a single layer and up to 5.6x on the whole networks.

cs.DC↗

arXiv2016

Person Re-identification: Past, Present and Future

Person re-identification (re-ID) has become increasingly popular in the community due to its application and research significance. It aims at spotting a person of interest in other cameras. In the early days, hand-crafted algorithms and small-scale evaluation were predominantly reported. Recent years have witnessed the emergence of large-scale datasets and deep learning systems which make use of large data volumes. Considering different tasks, we classify most current re-ID methods into two classes, i.e., image-based and video-based; in both tasks, hand-crafted and deep learning systems will be reviewed. Moreover, two new re-ID tasks which are much closer to real-world applications are described and discussed, i.e., end-to-end re-ID and fast re-ID in very large galleries. This paper: 1) introduces the history of person re-ID and its relationship with image classification and instance retrieval; 2) surveys a broad selection of the hand-crafted systems and the large-scale methods in both image- and video-based re-ID; 3) describes critical future directions in end-to-end re-ID and fast retrieval in large galleries; and 4) finally briefs some important yet under-developed issues.

cs.CV↗

arXiv2016

The expanding light echoes from supernova 2014J in M82

We present the measurement of the size and surface brightness of the expanding light echoes from supernova (SN) 2014J in the nearby starburst galaxy M82. Hubble Space Telescope (HST) ACS/WFC images were taken ~277 and ~416 days (after the time of B-band maximum light) in the filters F475W, F606W, and F775W, each combined with the three polarizing filters: POL0V, POL60V, and POL120V. The two epochs' imaging reveals the time evolution of at least two major echoes. Three concentric bright regions between position angles (PA, 0^{\circ} from North, counterclockwise). 80^{\circ} ~ 170^{\circ} have projected radius of 0.60" on the sky on ~277 days and expanding to 0.75" on ~416 days, corresponding to scattering materials at a foreground distance of 222\pm37 pc. Another fainter but evident light echo extending over a wide range of PA has radii of 0.75" and 0.96" on ~277 and ~416 days. This corresponds to scattering material at a foreground distance of 367\pm61 pc. Multiple light echoes with S/N > 2.5 reside at smaller radii on ~277 days but become less significant on ~416 days indicating a complex structure of foreground interstellar medium (ISM). The light echo shows bluer color than predicted under a Rayleigh scattering case. We also found the light echo brightened from V_{echo}=21.68\pm0.07 on 2014 September 5, to V_{echo}=21.05\pm0.08 on 2014 November 6, suggesting an enhancement of echoing materials at different distances projected on to the plane of the sky.

astro-ph.SR↗

arXiv2016

Inverse Magnetic Catalysis in the Soft-Wall Model of AdS/QCD

Magnetic effects on chiral phase transition have been investigated in a modified soft-wall AdS/QCD model, in which the dilaton field is taken to be negative at the ultraviolet region and positive at the infrared region as in Phys.Rev.D93(2016),101901 and JHEP1604(2016)036. The magnetic field is introduced into the background geometry by solving the Einstein-Maxwell system. After embedding the magnetized background geometry into the modified soft-wall model, the magnetic field dependent behavior of chiral condensate is worked out numerically. It is found that, in the chiral limit, the chiral phase transition remains as a second order at finite magnetic field $B$, while the symmetry restoration temperature and chiral condensate decrease with the increasing of magnetic field in small $B$ region. When including finite quark mass effect, the phase transition turns to be a crossover one, and the transition temperature still decreases with increasing magnetic field $B$ when $B$ is not very large. In this sense, inverse magnetic catalysis effect is observed in this modified soft-wall AdS/QCD model.

hep-th↗

arXiv2016

S-MART: Novel Tree-based Structured Learning Algorithms Applied to Tweet Entity Linking

Non-linear models recently receive a lot of attention as people are starting to discover the power of statistical and embedding features. However, tree-based models are seldom studied in the context of structured learning despite their recent success on various classification and ranking tasks. In this paper, we propose S-MART, a tree-based structured learning framework based on multiple additive regression trees. S-MART is especially suitable for handling tasks with dense features, and can be used to learn many different structures under various loss functions. We apply S-MART to the task of tweet entity linking --- a core component of tweet information extraction, which aims to identify and link name mentions to entities in a knowledge base. A novel inference algorithm is proposed to handle the special structure of the task. The experimental results show that S-MART significantly outperforms state-of-the-art tweet entity linking systems.

cs.CL↗