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:20230103T035307Z
LOCATION:Room 324\, Level 3\, West Wing
DTSTART;TZID=Asia/Seoul:20221206T140000
DTEND;TZID=Asia/Seoul:20221206T153000
UID:siggraphasia_SIGGRAPH Asia 2022_sess154_papers_234@linklings.com
SUMMARY:Motion In-betweening via Two-stage Transformers
DESCRIPTION:Technical Communications, Technical Papers\n\nMotion In-betwee
 ning via Two-stage Transformers\n\nQin, Zheng, Zhou\n\nWe present a deep l
 earning-based framework to synthesize motion in-betweening in a two-stage 
 manner. Given some context frames and a target frame, the system can gener
 ate plausible transitions with variable lengths in a non-autoregressive fa
 shion. The framework consists of two Transformer Encoder-based networks op
 erating in two stages: in the first stage a Context Transformer is designe
 d to generate rough transitions based on the context and in the second sta
 ge a Detail Transformer is employed to refine motion details. Compared to 
 existing Transformer-based methods which either use a complete Transformer
  Encoder-Decoder architecture or additional 1D convolutions to generate mo
 tion transitions, our framework achieves superior performance with less tr
 ainable parameters by only leveraging the Transformer Encoder and masked s
 elf-attention mechanism. To enhance the generalization of our transformer-
 based framework, we further introduce Keyframe Positional Encoding and Lea
 rned Relative Positional Encoding to make our method robust in synthesizin
 g longer transitions exceeding the maximum transition length during traini
 ng. Our framework is also artist-friendly by supporting full and partial p
 ose constraints within the transition, giving artists fine control over th
 e synthesized results. We benchmark our framework on the LAFAN1 dataset, a
 nd experiments show that our method outperforms the current state-of-the-a
 rt methods at a large margin (an average of 16% for normal-length sequence
 s and 55% for excessive-length sequences). Our method trains faster than t
 he RNN-based method and achieves a four-time speedup during inference. We 
 implement our framework into a production-ready tool inside an animation a
 uthoring software and conduct a pilot study to validate the practical valu
 e of our method.\n\nRegistration Category: FULL ACCESS, ON-DEMAND ACCESS\n
 \nLanguage: ENGLISH\n\nFormat: IN-PERSON, ON-DEMAND
URL:https://sa2022.siggraph.org/en/full-program/?id=papers_234&sess=sess15
 4
END:VEVENT
END:VCALENDAR