🧠 Daily Study Log [2025-07-23]
A day packed with project ideation, competition modeling, paper study, implementation, and GitHub cleanup.
Gradually converging toward more structured motion evaluation experiments.

💡 Project Idea — DanceMap: Mapping Genre-Based Dancer Hotspots in Seoul

Outlined a project concept to visualize where dancers gather by genre (e.g., hip-hop, waacking, popping) using online behavioral signals.
This idea connects my interest in pose estimation and urban culture data.

Key Points:

Aggregate social signals (hashtags, blogs, vlogs) to map genre-specific dance hubs
Help newcomers discover local dance scenes
Leverage NLP, geocoding, and spatiotemporal clustering
Visualize hotspots via Folium or Kakao Maps
Focus on cultural equity and community visibility

🗃 SQL Study

Reviewed subqueries with focus on:

SELECT-based nested queries
WHERE conditions using IN and EXISTS
Correlated vs. non-correlated subqueries
Applied to practice examples for filtering and aggregation

🏆 Competition — Electricity Usage Forecasting (DACON)

Logged three new submissions focused on fast Optuna tuning + feature engineering:

No.	Strategy	Notes	SMAPE
10	Ensemble + 4 new features + Optuna (n=15)	`hour_sin`, `cooling × hour`, etc. → improved performance	11.34327
11	XGBoost only + 4 features + Optuna (n=10)	Efficient single-model tuning	11.68912
12	XGBoost + 5 more engineered features	Slight overfitting observed despite additional features	13.66468

Next Steps:

Try StackingRegressor
Segment models by building type (e.g., hospitals vs. offices)
Experiment with lag features and rolling statistics

📄 Paper Review — Explainable Skeletal Evaluation for Martial Arts

Read:
“Explainable Quality Assessment of Effective Aligned Skeletal Representations for Martial Arts Movements” (Scientific Reports, 2025)

Highlights:

Combines Procrustes + DTW for spatial-temporal alignment
Uses Decision Tree, Logistic Regression, LSTM for motion scoring
Applies SHAP for explainable output per joint or body part
Strong candidate framework for practical feedback systems in sports/dance/rehab

🕺 CV Project — PoseSyncEvaluator: 2D Dance Alignment with Pose & DTW

Started implementing a pose-based movement evaluation system inspired by HDVR (ICCV 2021).
Goal: evaluate dance similarity using 2D pose sequences only.

Key Features:

Pose extraction: MediaPipe (33 keypoints)
Similarity metrics: cosine similarity + FastDTW
Visualizations: similarity curve, DTW path
Stored in 01_pose-sync-evaluator.ipynb (Colab)

This is the first step toward a larger cv-projects pipeline for real-time pose-based feedback.

🤝 Study Group & GitHub Maintenance

Shared 국민체조-based pose estimation experiment
Discussed possible generalization to healthcare applications
Compared related papers on skeleton alignment and scoring
Refactored repo structure, .gitignore, and notebooks for clarity

✅ TL;DR

📍 Idea: DanceMap — genre-wise dancer mapping via online signals
📍 SQL: Reviewed subquery logic and filtering techniques
📍 Competition: 3 new runs with Optuna + feature engineering (best SMAPE: 11.34)
📍 Paper: Read explainable skeleton evaluation paper using SHAP + DTW
📍 Project: Started PoseSyncEvaluator for 2D pose alignment
📍 Group: Shared code and aligned next steps with team
📍 GitHub: Updated repo structure and added notebooks