WEBVTT

00:00:00.500 --> 00:00:04.200
Hello, my name is Kyunggeun Roh from KAIST, 
and as written in the title 

00:00:04.201 --> 00:00:10.300
we preprocess Mediapipe keypoints with keypoint reconstruction 
and anchors for Isolated Sign Language Recognition (ISLR).

3
00:00:11.001 --> 00:00:18.300
In this research, we propose three different preprocessing methods 
to improve the performance of isolated sign language recognition.	

4
00:00:18.301 --> 00:00:23.500
First, we normalize hands separately based on anchor points 
to recognize hands better.

5
00:00:23.801 --> 00:00:31.000
Second, we reconstruct hands with bilinear interpolation 
for empty hand keypoints that Mediapipe fails to detect.

6
00:00:31.201 --> 00:00:38.200
Third, we standardize the input format by extending the length
of input sequences with a unified distribution.

7
00:00:38.501 --> 00:00:44.800
By integrating these methods, we significantly improve 
the performance and discuss about the results 

8
00:00:44.801 --> 00:00:49.300
on isolated sign language recognition datasets 
with transformer models.

9
00:00:49.450 --> 00:00:54.000
Details are shared in the poster and paper 
and thank you for listening.