Article Per Year (5 Year)
p-Index From 2021 - 2026
0.23
P-Index
This Author published in this journals
All Journal Informatika Mulawarman: Jurnal Ilmiah Ilmu Komputer
Kaleb, Viona Zatil Aqmar
Unknown Affiliation
Computer Science & IT

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

 1 
Efficiency of Temporal Convolutional Networks in Karate Kata Evaluation: A Comparative Study Kaleb, Viona Zatil Aqmar; Utomo, Wiranto Herry
Informatika Mulawarman : Jurnal Ilmiah Ilmu Komputer Vol 21, No 1 (2026): Informatika Mulawarman : Jurnal Ilmiah Ilmu Komputer
Publisher : Mulawarman University

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30872/jim.v21i1.28888

Abstract

The objective quantification of complex martial-arts movement is a long-standing challenge in computer vision. This study reports a feasibility prototype that pairs MoveNet Lightning pose estimation with a lightweight Temporal Convolutional Network for two tasks on a small custom dataset of four foundational Karate Kata (Heian Shodan through Heian Yondan): (1) multi-class kata recognition and (2) binary correctness evaluation. A 132-dimensional kinematic feature vector is constructed per frame from 17 MoveNet keypoints, combining hip-centered normalized coordinates, confidence scores, joint angles, bone-length ratios, and end-effector velocities. The full dataset comprises 100 lateral-view videos (80 train/10 validation/10 test) collected from a single Indonesian dojo. A multi-task model with two causal-dilated 1D-convolution layers (32 filters, dilations 1 and 2) is trained for up to 100 epochs on Apple M1 hardware over five random seeds. The kata head reaches 100% validation accuracy with zero variance across all five seeds; however, this result is interpreted with strong reservations because the validation set contains only 10 samples. The correctness head consistently collapses to majority-class prediction (60.0% ± 0.0%, identical to the all-positive baseline of 6/10). The contribution of this paper is therefore a fully reproducible end-to-end pipeline (MoveNet → 132-D features → multi-task TCN → real-time webcam demo) together with a candid characterization of the dataset-driven limits that block the correctness task at this scale.
Co-Authors Wiranto Herry Utomo