arXiv subjects

Jianming Zhang

Publications and source records attributed to Jianming Zhang.

96 records · Page 6Linked to original sources

arXiv2016

Top-down Neural Attention by Excitation Backprop

We aim to model the top-down attention of a Convolutional Neural Network (CNN) classifier for generating task-specific attention maps. Inspired by a top-down human visual attention model, we propose a new backpropagation scheme, called Excitation Backprop, to pass along top-down signals downwards in the network hierarchy via a probabilistic Winner-Take-All process. Furthermore, we introduce the concept of contrastive attention to make the top-down attention maps more discriminative. In experiments, we demonstrate the accuracy and generalizability of our method in weakly supervised localization tasks on the MS COCO, PASCAL VOC07 and ImageNet datasets. The usefulness of our method is further validated in the text-to-region association task. On the Flickr30k Entities dataset, we achieve promising performance in phrase localization by leveraging the top-down attention of a CNN model that has been trained on weakly labeled web images.

cs.CV↗

arXiv2016

Salient Object Subitizing

We study the problem of Salient Object Subitizing, i.e. predicting the existence and the number of salient objects in an image using holistic cues. This task is inspired by the ability of people to quickly and accurately identify the number of items within the subitizing range (1-4). To this end, we present a salient object subitizing image dataset of about 14K everyday images which are annotated using an online crowdsourcing marketplace. We show that using an end-to-end trained Convolutional Neural Network (CNN) model, we achieve prediction accuracy comparable to human performance in identifying images with zero or one salient object. For images with multiple salient objects, our model also provides significantly better than chance performance without requiring any localization process. Moreover, we propose a method to improve the training of the CNN subitizing model by leveraging synthetic images. In experiments, we demonstrate the accuracy and generalizability of our CNN subitizing model and its applications in salient object detection and image retrieval.

cs.CV↗

arXiv2015

Do Less and Achieve More: Training CNNs for Action Recognition Utilizing Action Images from the Web

Recently, attempts have been made to collect millions of videos to train CNN models for action recognition in videos. However, curating such large-scale video datasets requires immense human labor, and training CNNs on millions of videos demands huge computational resources. In contrast, collecting action images from the Web is much easier and training on images requires much less computation. In addition, labeled web images tend to contain discriminative action poses, which highlight discriminative portions of a video's temporal progression. We explore the question of whether we can utilize web action images to train better CNN models for action recognition in videos. We collect 23.8K manually filtered images from the Web that depict the 101 actions in the UCF101 action video dataset. We show that by utilizing web action images along with videos in training, significant performance boosts of CNN models can be achieved. We then investigate the scalability of the process by leveraging crawled web images (unfiltered) for UCF101 and ActivityNet. We replace 16.2M video frames by 393K unfiltered images and get comparable performance.

cs.CV↗

arXiv1995

Another Look at ``soft'' lepton pair production in nucleon-nucleon bremsstrahlung

We compare different formalisms for the calculation of lepton pair emission in hadron-hadron collisions and discuss the consequences of the approximations inherent to each of them. Using a Lorentz-covariant and gauge-invariant formalism proposed by Lichard, we calculate lepton pair emission via bremsstrahlung in proton-proton and neutron-proton reactions at energies between 1 and 5 GeV to leading order in the virtual photon four-momentum. These new results suggest that some previous bremsstrahlung calculations based on varieties of the soft-photon approximation might have somewhat overestimated the DLS low-mass dilepton cross sections. We find that the respective intensities of dilepton production through $p p$ and $n p$ bremsstrahlung are energy-dependent and become comparable at $E_{\rm kin} \agt $ 2 GeV. We also calculate the polar angle anisotropy of the lepton spectrum.

nucl-th↗

arXiv1994

Fragment Flow and the Nuclear Equation of State

We use the Boltzmann-Uehling-Uhlenbeck model with a momentum-dependent nuclear mean field to simulate the dynamical evolution of heavy ion collisions. We re-examine the azimuthal anisotropy observable, proposed as sensitive to the equation of state of nuclear matter. We obtain that this sensitivity is maximal when the azimuthal anisotropy is calculated for nuclear composite fragments, in agreement with some previous calculations. As a test case we concentrate on semi-central $^{197}{\rm Au}\ +\ ^{197}{\rm Au}$ collisions at 400 $A$ MeV.

nucl-th↗

arXiv1994

Momentum--dependent nuclear mean fields and collective flow in heavy ion collisions

We use the Boltzmann-Uehling-Uhlenbeck model to simulate the dynamical evolution of heavy ion collisions and to compare the effects of two parametrizations of the momentum--dependent nuclear mean field that have identical properties in cold nuclear matter. We compare with recent data on nuclear flow, as characterized by transverse momentum distributions and flow ($F$) variables for symmetric and asymmetric systems. We find that the precise functional dependence of the nuclear mean field on the particle momentum is important. With our approach, we also confirm that the difference between symmetric and asymmetric systems can be used to pin down the density and momentum dependence of the nuclear self consistent one--body potential, independently. All the data can be reproduced very well with a momentum--dependent interaction with compressibility K = 210 MeV.

nucl-th↗