# 9/26/24 Meeting Notes #6
# Generative Module
## Papers
- Customized Cloth Texture Generation Pipeline for 3D Virtual Try-On (2024)
- Garment3DGen: 3D Garment Stylization and Texture Generation (2024)
- Details refer to https://hackmd.io/@emps-113up/Hyyg6sa6A

# Classification Module
## Papers:

- Details refer to https://hackmd.io/@emps-113up/HJ3f2oTTA

# Work Flow using 2 classifcation model and 1 generative model
## Approach 1 方法一
- **Initial item classification 初始項目分類:**
Initially, the database will be empty. The first classification model will process items added by the user and classify each item into a suitable category (e.g., shirt, pants, etc.), extracting relevant attributes for each item.
最初，資料庫將是空的。第一個分類模型會處理使用者新增的項目，並將每個項目分類到合適的類別（例如，襯衫、褲子等），同時提取每個項目的相關屬性。
The database will have the item (an image), its category, and its attributes. 
資料庫將儲存該項目（圖像）、其類別及其屬性。

- **Style selection 風格選擇:**
After the user picks a desired style (e.g., "Punk"), the system selects items that match the style and puts them into a list.
當使用者選擇所需風格（例如，“龐克”風格）後，系統會選擇符合該風格的項目並將其放入列表中。
`For example, for "Punk" style, the system may pick items like a black leather jacket and black long pants. 對於 “龐克” 風格，系統可能會選擇黑色皮夾克和黑色長褲等項目。`

- **Creating combinations 創建組合：**
Using the list of selected items, the system generates various combinations of clothing.
系統會使用選定項目列表生成多種衣物組合。

- **Second classification model 第二分類模型：**
Each combination of clothes will be evaluated by the second classification model. This model, based on the paper `Item-region-based style classification network (IRSN)`, extracts features the overall look as well as individual items.
每個衣物組合都會通過第二個分類模型進行評估。該模型基於`Item-region-based style classification network (IRSN)` 這篇論文，提取整體外觀和單個項目的特徵。
`For example, while a black shirt and black pants may individually be considered "punk," their combination might be classified as "casual." 雖然黑色襯衫和黑色褲子個別來看可能是“龐克”風格，但它們的組合可能會被分類為“休閒”風格。`

- **Combination evaluation 組合評估 :**
If the user has 3 shirts and 2 pants, the system will create 6 combinations. Each combination is processed by the second classification model to determine its accuracy rate for the desired style (e.g., 90% confidence in "punk" style, 60% not "punk").
如果使用者有3件襯衫和2條褲子，系統將生成6種組合。每個組合都會通過第二個分類模型進行處理，以確定其在所需風格中的準確率（例如，90%符合“龐克”風格，60%不符合“龐克”）。
-- If the style is not approved, the system moves on to the next combination.
如果風格不被批准，系統將繼續處理下一個組合。
-- If the style is approved, the combination will be added to a temporary list of suitable 2D combinations that match the desired style.
如果風格被批准，該組合將被添加到符合所需風格的臨時2D組合列表中。

- **Creating a list of suitable combinations 創建合適組合的列表：**
The output of this evaluation process will be a list containing all 2D combinations that have been verified by the classification model as suitable for the selected style.
此評估過程的輸出將是一個包含所有經過分類模型驗證的適合所選風格的2D組合列表。

- **Generative model processing 生成模型處理:**
The items in the temporary list will be processed through phase I and phase II of the generative model, one by one.
臨時列表中的項目將逐一通過生成模型的第一階段和第二階段進行處理。

- **User interface and selection 使用者界面與選擇:**
After generating 3D models for all suitable combinations, the system presents the user with a list of 2D previews. The user can select which combination they want to visualize.
在為所有合適的組合生成3D模型後，系統將向使用者展示2D預覽列表，使用者可以選擇他們想要視覺化的組合。

- **Final output 最終輸出:**
The selected outfit is generated as a 3D model, ready to be visualized on the user's body.
所選的服裝將生成一個3D模型，準備在使用者的身體上進行視覺化。

### Flow Diagram
**User interface:**
**Approach**

## Approach 2 方式二
- **Initial item classification 初始項目分類：**
Initially database will be empty. The first classification model will process items added by user and classifies each item into a suitable category (e.g., shirt, pants, etc.)
最初資料庫將是空的。第一個分類模型會處理使用者新增的項目，並將每個項目分類到合適的類別（例如，襯衫、褲子等）。

- **Combination process 組合過程:**
When enough items are added (at least one shirt and one pair of pants), the system generates a combination and sends it to the second classification model. If there aren't enough items, the system waits for more to be added.
當新增足夠的項目（至少一件襯衫和一條褲子）後，系統將生成一個組合，並將其發送到第二個分類模型。如果項目數不足，系統將等待更多項目被新增。

- **Second classification model 第二分類模型：**
This model evaluates the outfit combination and assigns an accuracy rate and style.
該模型會評估衣物組合，並分配一個準確率和風格。
-- If the accuracy rate is below the approval threshold (e.g., less than 80% for a "punk" style), the outfit is not included in our defined styles.
如果準確率低於批准閾值（例如，低於80%的“龐克”風格），該組合不會被包括在我們定義的風格中。
-- If the accuracy is above the threshold, the combination and its style are saved in the database, with each item saved individually (image by image).
如果準確率高於閾值，該組合及其風格將被儲存在資料庫中，並將每個項目單獨儲存（逐一圖像存儲）。

- **User requests outfit 使用者請求服裝組合：**
When the user requests an outfit in a specific style, the system retrieves relevant outfits from the database and processes them through phase I and phase II of the generative model.
當使用者請求特定風格的服裝時，系統將從資料庫中檢索相關的服裝，並將其通過生成模型的第一階段和第二階段進行處理。

- **User selection 使用者選擇：**
The system generates 3D models (from phase II) for all suitable combinations and presents 2D previews to the user. The user selects their desired combination for final generation.
系統會為所有合適的組合生成3D模型（來自第二階段），並向使用者展示2D預覽。使用者可以選擇他們想要生成的最終組合。

- **Final output 最終輸出：**
The selected outfit is generated as a 3D model, ready to be visualized on the user's body.
所選的服裝將生成一個3D模型，準備在使用者的身體上進行視覺化。

### Flow Diagram

# To do for next meeting:
- Classification:
1. Compare all papers to see which one is the best to implement (consider: easy to code, good performance, good result)

---
Previous: [9/19 Meeting Notes #5](https://hackmd.io/@emps-113up/meeting5)
Next: [10/3 Meeting Notes #7](https://hackmd.io/@emps-113up/meeting7)
Full Content List [here](https://hackmd.io/@emps-113up/full-list)