Meta tags:
description=;
author=;

Headings (most frequently used words):

joint, to, mesh, learning, of, 3d, shape, retrieval, and, deformation, abstract, video, materials, our, approach, image, point, cloud, training, on, neural, cages, automatic, segmentation, applications, citation, acknowledgements,

Text of the page (most frequently used words):
the (40), and (24), deformation (17), our (17), joint (12), that (10), shape (9), #retrieval (8), models (7), part (6), also (6), source (6), target (6), module (6), learning (5), match (5), from (4), such (4), use (4), training (4), novel (4), neural (4), aware (4), results (4), mesh (4), image (4), space (4), show (3), method (3), modules (3), chair (3), point (3), cloud (3), network (3), given (3), database (3), this (2), work (2), adobe (2), mikaela (2), angelina (2), vladimir (2), kim (2), minhyuk (2), sung (2), noam (2), aigerman (2), siddhartha (2), chaudhuri (2), leonidas (2), guibas (2), conference (2), computer (2), vision (2), pattern (2), recognition (2), cvpr (2), 2021 (2), applications (2), databases (2), are (2), approach (2), not (2), function (2), but (2), other (2), state (2), art (2), cages (2), retrieved (2), approaches (2), input (2), sofa (2), structure (2), thickness (2), visualization (2), qualitative (2), setup (2), achieves (2), best (2), shown (2), table (2), legs (2), based (2), per (2), propose (2), its (2), with (2), used (2), learn (2), embedding (2), more (2), pairs (2), demonstrate (2), jointly (2), outperforms (2), non (2), baselines (2), supported, grant, samsung, gro, program, vannevar, bush, faculty, fellowship, gifts, autodesk, snap, acknowledgements, inproceedings, cvpr21, title, author, booktitle, ieee, year, citation, pipeline, fitting, product, images, google, search, real, world, scans, manual, fine, grain, annotating, tedious, task, large, scarce, works, automatically, segmented, those, found, complementme, enabling, wider, array, automatic, segmentation, advantages, specific, improves, observe, better, geometry, moreover, able, local, geometries, height, seat, base, cabinet, without, shelves, back, during, deformable, sources, retrieve, subset, their, proximity, right, fit, each, uses, encoded, global, codes, predict, parameters, code, poster, slides, paper, materials, video, technique, for, producing, high, quality, object, scan, retrieving, existing, then, deforming, parts, unlike, previous, dependently, focus, either, procedure, simultaneously, trains, along, embed, ding, enables, amenable, tar, get, after, appropriate, fact, guide, train, invests, capacity, ing, deformations, between, meaningful, further, can, inconsistent, diverse, structures, shapes, benefits, only, framework, proposed, recent, years, lastly, trained, various, abstract, learned, find, model, heterogeneous, align, static, stanford, university, research, kaist, iitb,

Text of the page (random words):
joint learning of retrieval and deformation joint learning of 3d shape retrieval and deformation mikaela angelina uy 1 vladimir g kim 2 minhyuk sung 3 noam aigerman 2 siddhartha chaudhuri 2 4 leonidas guibas 1 1 stanford university 2 adobe research 3 kaist 4 iitb conference on computer vision and pattern recognition cvpr 2021 given an input target we use jointly learned retrieval and deformation modules to find a source model in a heterogeneous database and align it to the target we demonstrate that our joint learning outperforms static retrieval and non joint baselines abstract we propose a novel technique for producing high quality 3d models that match a given target object image or scan our method is based on retrieving an existing shape from a database of 3d models and then deforming its parts to match the target shape unlike previous approaches that in dependently focus on either shape retrieval or deformation we propose a joint learning procedure that simultaneously trains the neural deformation module along with the embed ding space used by the retrieval module this enables our network to learn a deformation aware embedding space so that retrieved models are more amenable to match the tar get after an appropriate deformation in fact we use the embedding space to guide the shape pairs used to train the deformation module so that it invests its capacity in learn ing deformations between meaningful shape pairs further more our novel part aware deformation module can work with inconsistent and diverse part structures on the source shapes we demonstrate the benefits of our joint training not only on our novel framework but also on other state of the art neural deformation modules proposed in recent years lastly we also show that our jointly trained method outperforms various non joint baselines video materials paper slides poster code our approach during training given a target image or a point cloud and a database of deformable sources we retrieve a subset of source models based on their proximity in the retrieval space and use the structure aware deformation module right to fit each source our deformation module uses encoded target global and per part source codes to predict per part deformation parameters image to mesh visualization of qualitative results on the image to mesh setup our network achieves the best results as shown by the thickness of legs of the chair the shape of the back and the shape of the legs of the table point cloud to mesh visualization of qualitative results on the point cloud to mesh setup our network also achieves the best results as shown by the height of the chair seat the base of the table and the cabinet without shelves joint training on neural cages we also show that the advantages of joint training is not specific to our novel deformation function but also improves other state of the art deformation modules such as neural cages observe that the retrieved models of our joint approaches better match the geometry of the input chair and sofa moreover our structure aware deformation function is able to match the local part geometries such as the thickness of the sofa automatic segmentation as manual fine grain part annotating is a tedious task large databases of such models are scarce we also show that our approach works on automatically segmented models such as those found in complementme enabling the use of our method on a wider array of source databases applications applications of our pipeline on fitting to product images from google search and to real world scans citation inproceedings uy joint cvpr21 title joint learning of 3d shape retrieval and deformation author mikaela angelina uy and vladimir g kim and minhyuk sung and noam aigerman and siddhartha chaudhuri and leonidas guibas booktitle ieee conference on computer vision and pattern recognition cvpr year 2021 acknowledgements this work is supported by a grant from the samsung gro program a vannevar bush faculty fellowship and gifts from adobe autodesk and snap