BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Seoul
X-LIC-LOCATION:Asia/Seoul
BEGIN:STANDARD
TZOFFSETFROM:+0900
TZOFFSETTO:+0900
TZNAME:KST
DTSTART:18871231T000000
DTSTART:19881009T020000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20230103T035312Z
LOCATION:Room 325-AB\, Level 3\, West Wing
DTSTART;TZID=Asia/Seoul:20221209T140000
DTEND;TZID=Asia/Seoul:20221209T153000
UID:siggraphasia_SIGGRAPH Asia 2022_sess178_papers_290@linklings.com
SUMMARY:NeuralMarker: A Framework for Learning General Marker Corresponden
 ce
DESCRIPTION:Technical Communications, Technical Papers\n\nNeuralMarker: A 
 Framework for Learning General Marker Correspondence\n\nHuang, Pan, Pan, B
 ian, Xu...\n\nWe tackle the problem of estimating correspondences from a g
 eneral marker, such as a movie poster, to an image that captures such a ma
 rker. Conventionally, this problem is addressed by fitting a homography mo
 del based on sparse feature matching. However, they are only able to handl
 e plane-like markers and the sparse features do not sufficiently utilize a
 ppearance information. In this paper, we propose a novel framework NeuralM
 arker, training a neural network estimating dense marker correspondences u
 nder various challenging conditions, such as marker deformation, harsh lig
 hting, etc. Deep learning has presented an excellent performance in corres
 pondence learning once provided with sufficient training data. However, an
 notating pixel-wise dense correspondence for training marker correspondenc
 e is too expensive. We observe that the challenges of marker correspondenc
 e estimation come from two individual aspects: geometry variation and appe
 arance variation. We, therefore, design two components addressing these tw
 o challenges in NeuralMarker. First, we create a synthetic dataset FlyingM
 arkers containing marker-image pairs with ground truth dense correspondenc
 es. By training with FlyingMarkers, the neural network is encouraged to ca
 pture various marker motions. Second, we propose the novel Symmetric Epipo
 lar Distance (SED) loss, which enables learning dense correspondence from 
 posed images. Learning with the SED loss and the cross-lighting posed imag
 es collected by Structure-from-Motion (SfM), NeuralMarker is remarkably ro
 bust in harsh lighting environments and avoids the synthetic image bias. B
 esides, we also propose a novel marker correspondence evaluation method ci
 rcumstancing annotations on real marker-image pairs and create a new bench
 mark. We show that NeuralMarker significantly outperforms previous methods
  and enables new interesting applications, including Augmented Reality (AR
 ) and video editing.\n\nRegistration Category: FULL ACCESS, ON-DEMAND ACCE
 SS\n\nLanguage: ENGLISH\n\nFormat: IN-PERSON, ON-DEMAND
URL:https://sa2022.siggraph.org/en/full-program/?id=papers_290&sess=sess17
 8
END:VEVENT
END:VCALENDAR