2026.09.30

Participants: Hyeon U Jeong, Dongkun Moon, Kihwan Kim, Juhyeon Choi

Discussion

Clothing Analysis

• Reviewed the current clothing analysis model and measurement accuracy.
• Agreed to run additional tests to evaluate measurement errors.
• If automatic measurement is not sufficiently accurate, manual user input may be used as a fallback.

Database

• Discussed the initial database structure.
• Main entities will include User, Clothes, and Outfit.
• Detailed fields will be finalized after testing the clothing analysis model.

UI / Wireframe

• Reviewed the wireframe for the main screens and onboarding flow.
• The wireframe will be shared through Figma and refined based on the current implementation.
• Decided to keep StyleMate as the project name for now.

Next Feature: Outfit Recommendation

• Agreed to work on the outfit recommendation system as the next major feature.
• Recommendations will use clothes from the user's registered wardrobe and onboarding preferences.
• Users will be able to refine recommendations through conversational requests.

Iteration & Documentation

• Reviewed the progress of Iteration 1.
• Requirements and design documents will be revised according to the course guidelines.
• The team will begin the outfit recommendation feature after completing the current features.

Action Items

• Test clothing analysis model and measurement accuracy
• Finalize clothing attributes and DB schema
• Share and refine Figma wireframe
• Update Requirements & Design Documents
• Complete clothing registration flow
• Start designing the outfit recommendation feature

---

2026.10.07

Participants: Hyeon U Jeong, Dongkun Moon, Kihwan Kim, Juhyeon Choi

Discussion

Recommendation Approach

• Compared two candidate approaches for outfit recommendation:
  • Slot-based: group clothes into slots (top, bottom, outer, shoes; hat optional), fill the remaining slots around an item the user locks, and score combinations. Simple and certainly implementable, but limited for one-pieces and layered outfits.
  • Data-driven: learn which items go together from real outfit data using item embeddings. More general, but data collection and fashion edge cases make it costly.
• The slot-based method will serve as the baseline; the final approach will be decided early next week.
• Recommendations will consider the whole wardrobe, without rule-based pre-filtering.

Outfit Scoring

• Color will be the main scoring factor, as color harmony is easy to quantify and well documented.
• Fit will be derived by comparing garment measurements with the user's body measurements.
• Material and texture are excluded for now due to their complexity.

Personalization

• User feedback (likes/dislikes and chat requests) will adjust the scoring weights.
• Long-term preference learning may be added in a later iteration.

Data Sources

• Discussed crawling Musinsa outfit snaps; feasibility, time cost and terms of use need to be checked.
• Found the AI Hub K-Fashion dataset (~1.2M expert-labeled outfit photos with category, fit and style labels). It is mostly women's clothing; label details will be reviewed after download.

Teaching Staff Feedback

• Keep the search space small: one top, one bottom, shoes, and an optional hat.
• Focus on diverse recommendation features (e.g., locking an item, swapping parts, adapting to weather) rather than claiming better taste than other services.
• Evaluate with metrics such as task completion time, and consider latency if we serve our own model.
• Document the models used and the rationale behind them in the Design Documentation.

Action Items

• Test whether Musinsa snaps can be crawled
• Analyze how Acloset scores outfits
• Research outfit scoring methods
• Download and review the K-Fashion dataset
• Meet early next week to finalize the recommendation approach
• Add the models used and their rationale to the Design Documentation

---

2026.10.12

Participants: Hyeon U Jeong, Dongkun Moon, Kihwan Kim, Juhyeon Choi

Recommendation Algorithm and Dataset

• Dataset:
• Algorithm: use self-attention model

cf) find other dataset

• From Musinsa, KREAM, etc.
• Not essential, just for improving recommendation quality.

Personalization through LLM chatting

• Use GEMINI API KEY for natural-language conversation, receive recommendation request via formal prompt
• Or we can use other LLM models to save tokens

Design DB Schema and Onboarding UI

• Finalize DB Schema of 3 tables(user, clothes, outfit)
• Implement total user flow UI and test

Next Feature

• Clothes(not in my wardrobe) recommendation
• Clothing preview system
• Image preprocessing(remove backgrounds, ...) when saving clothes

Action Items

• Test Recommandation system works
• Find other dataset for model training
• Implement total user flow(login ~ recommendation) and finalize db schema
• Implement total UI
• Find other LLM models for chatting