-
Video Prediction Recalling Long-term Motion Context via Memory Alignment Learning
Authors:
Sangmin Lee,
Hak Gu Kim,
Dae Hwi Choi,
Hyung-Il Kim,
Yong Man Ro
Abstract:
Our work addresses long-term motion context issues for predicting future frames. To predict the future precisely, it is required to capture which long-term motion context (e.g., walking or running) the input motion (e.g., leg movement) belongs to. The bottlenecks arising when dealing with the long-term motion context are: (i) how to predict the long-term motion context naturally matching input seq…
▽ More
Our work addresses long-term motion context issues for predicting future frames. To predict the future precisely, it is required to capture which long-term motion context (e.g., walking or running) the input motion (e.g., leg movement) belongs to. The bottlenecks arising when dealing with the long-term motion context are: (i) how to predict the long-term motion context naturally matching input sequences with limited dynamics, (ii) how to predict the long-term motion context with high-dimensionality (e.g., complex motion). To address the issues, we propose novel motion context-aware video prediction. To solve the bottleneck (i), we introduce a long-term motion context memory (LMC-Memory) with memory alignment learning. The proposed memory alignment learning enables to store long-term motion contexts into the memory and to match them with sequences including limited dynamics. As a result, the long-term context can be recalled from the limited input sequence. In addition, to resolve the bottleneck (ii), we propose memory query decomposition to store local motion context (i.e., low-dimensional dynamics) and recall the suitable local context for each local part of the input individually. It enables to boost the alignment effects of the memory. Experimental results show that the proposed method outperforms other sophisticated RNN-based methods, especially in long-term condition. Further, we validate the effectiveness of the proposed network designs by conducting ablation studies and memory feature analysis. The source code of this work is available.
△ Less
Submitted 2 April, 2021;
originally announced April 2021.
-
Neutron beam test of CsI crystal for dark matter search
Authors:
H. Park,
D. H. Choi,
I. S. Hahn,
M. J. Hwang,
W. G. Kang,
H. J. Kim,
J. H. Kim,
S. C. Kim,
S. K. Kim,
T. Y. Kim,
Y. D. Kim,
Y. J. Kwon,
H. S. Lee,
J. H. Lee,
M. H. Lee,
S. E. Lee,
S. H. Noh,
I. H. Park,
E. S. Seo,
E. Won,
H. Y. Yang,
M. S. Yang,
I. Yu
Abstract:
We have studied the response of Tl-doped and Na-doped CsI crystals to nuclear recoils and $γ$'s below 10 keV. The response of CsI crystals to nuclear recoil was studied with mono-energetic neutrons produced by the $^3$H(p,n)$^3$He reaction. This was compared to the response to Compton electrons scattered by 662 keV $γ$-ray. Pulse shape discrimination between the response to these $γ$'s and nucle…
▽ More
We have studied the response of Tl-doped and Na-doped CsI crystals to nuclear recoils and $γ$'s below 10 keV. The response of CsI crystals to nuclear recoil was studied with mono-energetic neutrons produced by the $^3$H(p,n)$^3$He reaction. This was compared to the response to Compton electrons scattered by 662 keV $γ$-ray. Pulse shape discrimination between the response to these $γ$'s and nuclear recoils was studied, and quality factors were estimated. The quenching factors for nuclear recoils were derived for both CsI(Na) and CsI(Tl) crystals.
△ Less
Submitted 13 February, 2002;
originally announced February 2002.
-
Parameterizations of 1-bridge torus knots
Authors:
Doo Ho Choi,
Ki Hyoung Ko
Abstract:
A 1-bridge torus knot in a 3-manifold of genus $\le 1$ is a knot drawn on a Heegaard torus with one bridge. We give two types of normal forms to parameterize the family of 1-bridge torus knots that are similar to the Schubert's normal form and the Conway's normal form for 2-bridge knots. For a given Schubert's normal form we give algorithms to determine the number of components and to compute th…
▽ More
A 1-bridge torus knot in a 3-manifold of genus $\le 1$ is a knot drawn on a Heegaard torus with one bridge. We give two types of normal forms to parameterize the family of 1-bridge torus knots that are similar to the Schubert's normal form and the Conway's normal form for 2-bridge knots. For a given Schubert's normal form we give algorithms to determine the number of components and to compute the fundamental group of the complement when the normal form determines a knot. We also give a description of the double branched cover of an ambient 3-manifold branched along a 1-bridge torus knot by using its Conway's normal form and obtain an explicit formula for the first homology of the double cover.
△ Less
Submitted 2 January, 2002; v1 submitted 11 December, 2001;
originally announced December 2001.
