BEGIN:VCALENDAR VERSION:2.0 PRODID:-//jEvents 2.0 for Joomla//EN CALSCALE:GREGORIAN METHOD:PUBLISH BEGIN:VEVENT UID:86svfdcq7djbnn7884ndscbhi8@google.com CATEGORIES:RTportal.ru CREATED:20170606T092330 SUMMARY::christmas_tree: Christmas Colloquium on Computer Vision LOCATION:Scolkovo, Moscow Oblast, Russia DESCRIPTION;ENCODING=QUOTED-PRINTABLE:Six researchers will present their recent works from ECCV’14, CVPR’15, ICCV ’15, ICML’15, ICLR’16(submitted) of this year. The language of the talk wil l be chosen by the speakers.

Center for Data-Intensive Science and Engineering (CDISE) of Skoltech invites you to participate (pre-registrati on needed)!
For the pre-registration, please send email with the subjec t ‘CCCV registration’ and your name to Адрес электронной почты защищен от спам-ботов. Для просмотра а дреса в вашем браузере должен быть включен Javascript. (CC Адрес электронной почты защищен от спам-ботов. Для просмотра адреса в вашем браузере должен быть включен J avascript.)

SCHEDULE:

14:00 — 14:30
Asya Pentina, IST Austria, Vienna
Title: Curriculum learning of multipl e tasks
Abstract: Sharing information between multiple tasks enables al gorithms to achieve good generalization performance even from small amounts of training data. However, in a realistic scenario of multi-task learning not all tasks are equally related to each other, hence it could be advantag eous to transfer information only between the most related tasks.
In th is work we propose an approach that processes multiple tasks in a sequence with sharing between subsequent tasks instead of solving all tasks jointly. Subsequently, we address the question of curriculum learning of tasks, i.e . finding the best order of tasks to be learned. Our approach is based on a generalisation bound criterion for choosing the task order that optimises the average expected classification performance over all tasks.Our experime ntal results show that learning multiple related tasks sequentially can be more effective than learning them jointly, the order in which tasks are bei ng solved affects the overall performance, and that our model is able to au tomatically discover a favourable order of tasks.

14:30 — 15:00
Sergey Zagoruyko, École de Ponts ParisTech (talk based on the work at Fac ebook AI Research)
Title: FAIRCNN MSCOCO object detection/segmentation challenge submission.
Abstract: Our approach is built on DeepMask propo sals fed into the Fast R-CNN pipeline. The DeepMask proposals have been sub stantially improved to encourage proposal diversity and mask quality. We al so augmented our CNN classifier with a novel foveal structure, skip-connect ions, an improved cost function that encourages better localization, and a few additional modifications. Finally we utilize ensembling (model and infe rence) to further improve performance.

15:00 — 15:15
Break

15:15 — 15:45
Anton Osokin, INRIA/École Normale Supérieure, Paris
Title: Context-aware CNNs for person head detection
Abstract: Pers on detection is a key problem for many computer vision tasks. While face de tection has reached maturity, detecting people under a full variation of ca mera view-points, human poses, lighting conditions and occlusions is still a difficult challenge. In this work we focus on detecting human heads in na tural scenes. Starting from the recent local R-CNN object detector, we exte nd it with two types of contextual cues. First, we leverage person-scene re lations and propose a Global CNN model trained to predict positions and sca les of heads directly from the full image. Second, we explicitly model pair wise relations among objects and train a Pairwise CNN model using a structu red-output surrogate loss. The Local, Global and Pairwise models are combin ed into a joint CNN framework. To train and test our full model, we introdu ce a large dataset composed of 369,846 human heads annotated in 224,740 mov ie frames. We evaluate our method and demonstrate improvements of person he ad detection against several recent baselines in three datasets. We also sh ow improvements of the detection speed provided by our model.

15:4 5 — 16:15
Danila Potapov, INRIA-LEAR, Grenoble
Title: Category-spec ific video summarization
Abstract: In large video collections with clus ters of typical categories, such as “birthday party” or “flash-mob”, catego ry-specific video summarization can produce higher quality video summaries than unsupervised approaches that are blind to the video category.
Give n a video from a known category, our approach first efficiently performs a temporal segmentation into semantically-consistent segments, delimited not only by shot boundaries but also general change points. Then, equipped with an SVM classifier, our approach assigns importance scores to each segment. The resulting video assembles the sequence of segments with the highest sc ores. The obtained video summary is therefore both short and highly informa tive. Experimental results on videos from the multimedia event detection (M ED) dataset of TRECVID’11 show that our approach produces video summaries w ith higher relevance than the state of the art.

16:15 — 16:30
Break

16:30 — 17:00
Michael Figurnov, Skoltech, Moscow
Tit le: PerforatedCNNs: Acceleration through Elimination of Redundant Convoluti ons
Abstract: We propose a novel approach to reduce the computational c ost of evaluation of convolutional neural networks, a factor that has hinde red their deployment in low-power devices such as mobile phones. Inspired b y the loop perforation technique from source code optimization, we speed up the bottleneck convolutional layers by skipping their evaluation in some o f the spatial positions. We propose and analyze several strategies of choos ing these positions. Our method allows to reduce the evaluation time of mod ern convolutional neural networks by 50% with a small decrease in accuracy. More details can be found in our ICLR 2016 submission: http://arxiv.org/pdf/1504.08362v2.pdf

17:00 — 17:30
Yaroslav Ganin, Skoltech, Moscow
Title: Unsuperv ised Domain Adaptation by Backpropagation
Abstract: Top-performing deep architectures are trained on massive amounts of labeled data. In the absen ce of labeled data for a certain task, domain adaptation often provides an attractive option given that labeled data of similar nature but from a diff erent domain (e.g. synthetic images) are available. Here, we propose a new approach to domain adaptation in deep architectures that can be trained on large amount of labeled data from the source domain and large amount of unl abeled data from the target domain (no labeled target-domain data is necess ary). As the training progresses, the approach promotes the emergence of “d eep” features that are (i) discriminative for the main learning task on the source domain and (ii) invariant with respect to the shift between the dom ains. We show that this adaptation behaviour can be achieved in almost any feed-forward model by augmenting it with few standard layers and a simple n ew gradient reversal layer. The resulting augmented architecture can be tra ined using standard backpropagation. Overall, the approach can be implement ed with little effort using any of the deep-learning packages. The method p erforms very well in a series of image classification experiments, achievin g adaptation effect in the presence of big domain shifts and outperforming previous state-of-the-art on Office datasets. We also validate the approach for descriptor learning task in the context of person re-identification ap plication.

http://sites.skoltech.ru/compvision/events/cccv2015/ This event was imported from: https://rtportal.ru/index.php/kalendar-sobytij/eventdet ail/680/-/christmas-tree-christmas-colloquium-on-computer-vision?tmpl=compo nent DTSTAMP:20260812T230909Z DTSTART;TZID=Europe/Moscow:20151228T140000 DTEND;TZID=Europe/Moscow:20151228T180000 SEQUENCE:0 TRANSP:OPAQUE END:VEVENT END:VCALENDAR