Spaces:
Running
Running
1. update README
Browse files2. adjust sample image location
3. adjust some display texts in app.py
- README.md +105 -1
- app.py +4 -5
- car.jpg → sample_images/car.jpg +0 -0
README.md
CHANGED
|
@@ -9,4 +9,108 @@ app_file: app.py
|
|
| 9 |
pinned: false
|
| 10 |
---
|
| 11 |
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
pinned: false
|
| 10 |
---
|
| 11 |
|
| 12 |
+
# 🎯 Text-Guided Image Segmentation Demo
|
| 13 |
+
|
| 14 |
+
基於 **Grounding DINO** 和 **SAM (Segment Anything Model)** 的文字引導圖片分割應用,使用 Gradio 構建互動式介面。
|
| 15 |
+
|
| 16 |
+
## 🚀 快速開始
|
| 17 |
+
|
| 18 |
+
### 安裝依賴
|
| 19 |
+
|
| 20 |
+
```bash
|
| 21 |
+
pip install -r requirements.txt
|
| 22 |
+
```
|
| 23 |
+
|
| 24 |
+
### 運行應用
|
| 25 |
+
|
| 26 |
+
```bash
|
| 27 |
+
gradio app.py
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
或
|
| 31 |
+
|
| 32 |
+
```bash
|
| 33 |
+
python app.py
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
應用將在 `http://localhost:7860` 啟動。
|
| 37 |
+
|
| 38 |
+
## 📦 依賴項
|
| 39 |
+
|
| 40 |
+
- `gradio` - 互動式 Web 介面
|
| 41 |
+
- `transformers` - Hugging Face 模型庫
|
| 42 |
+
- `torch` - PyTorch 深度學習框架
|
| 43 |
+
- `Pillow` - 圖像處理
|
| 44 |
+
- `numpy` - 數值計算
|
| 45 |
+
|
| 46 |
+
## 🎮 使用方法
|
| 47 |
+
|
| 48 |
+
1. **上傳圖片**:點擊或拖拽圖片到輸入區域
|
| 49 |
+
2. **輸入文字提示**:
|
| 50 |
+
- 單個物件:`car`、`person`、`sky`
|
| 51 |
+
- 多個物件:`car. sky. road.`(用句號分隔)
|
| 52 |
+
3. **點擊 Segment 按鈕**:開始分割
|
| 53 |
+
4. **查看結果**:
|
| 54 |
+
- 左側顯示分割遮罩(不同物件用不同顏色)
|
| 55 |
+
- 右側顯示除錯資訊(檢測數量、標籤、信心度)
|
| 56 |
+
|
| 57 |
+
## 🎨 顏色標示
|
| 58 |
+
|
| 59 |
+
檢測到的物件會按順序使用以下顏色作為mask:
|
| 60 |
+
|
| 61 |
+
1. 🔴 紅色
|
| 62 |
+
2. 🟢 綠色
|
| 63 |
+
3. 🔵 藍色
|
| 64 |
+
4. 🟡 黃色
|
| 65 |
+
|
| 66 |
+
## 🔧 技術架構
|
| 67 |
+
|
| 68 |
+
### 模型
|
| 69 |
+
|
| 70 |
+
- **Grounding DINO** (`IDEA-Research/grounding-dino-base`)
|
| 71 |
+
- 用於零樣本物件檢測
|
| 72 |
+
- 根據文字描述定位物件
|
| 73 |
+
|
| 74 |
+
- **SAM** (`facebook/sam-vit-base`)
|
| 75 |
+
- 用於精確分割
|
| 76 |
+
- 基於檢測框生成高質量遮罩
|
| 77 |
+
|
| 78 |
+
### 工作流程
|
| 79 |
+
|
| 80 |
+
```
|
| 81 |
+
輸入圖片 + 文字提示
|
| 82 |
+
↓
|
| 83 |
+
Grounding DINO 檢測物件
|
| 84 |
+
↓
|
| 85 |
+
SAM 生成分割遮罩
|
| 86 |
+
↓
|
| 87 |
+
多物件遮罩疊加(不同顏色)
|
| 88 |
+
↓
|
| 89 |
+
輸出結果
|
| 90 |
+
```
|
| 91 |
+
|
| 92 |
+
## 💡 使用技巧
|
| 93 |
+
|
| 94 |
+
### 提高檢測準確度
|
| 95 |
+
|
| 96 |
+
1. **使用具體描述**:`blue car` 比 `car` 更精確
|
| 97 |
+
2. **調整閾值**:當前閾值為 0.15,可在 source code 中調整
|
| 98 |
+
3. **多次嘗試**:嘗試不同的文字表達方式或是使用英文描述
|
| 99 |
+
|
| 100 |
+
## 📝 程式碼結構
|
| 101 |
+
|
| 102 |
+
```
|
| 103 |
+
text-image-seg/
|
| 104 |
+
├── app.py
|
| 105 |
+
├── requirements.txt
|
| 106 |
+
├── README.md
|
| 107 |
+
├── sample_images/
|
| 108 |
+
├── .gitattributes
|
| 109 |
+
└── .gitignore
|
| 110 |
+
```
|
| 111 |
+
|
| 112 |
+
## 🔗 相關資源
|
| 113 |
+
|
| 114 |
+
- [Grounding DINO](https://github.com/IDEA-Research/GroundingDINO)
|
| 115 |
+
- [Segment Anything Model (SAM)](https://github.com/facebookresearch/segment-anything)
|
| 116 |
+
- [Gradio 文檔](https://www.gradio.app/docs)
|
app.py
CHANGED
|
@@ -142,8 +142,8 @@ def segment_and_display(image, text):
|
|
| 142 |
|
| 143 |
result_image, debug_info = segment_image_with_text(image, text)
|
| 144 |
|
| 145 |
-
output_text = f"檢測到 {debug_info.get('num_detections', 0)} 個物件\n"
|
| 146 |
-
output_text += f"
|
| 147 |
|
| 148 |
if 'labels' in debug_info and len(debug_info['labels']) > 0:
|
| 149 |
output_text += "檢測結果:\n"
|
|
@@ -151,7 +151,6 @@ def segment_and_display(image, text):
|
|
| 151 |
color_map = ['紅色', '綠色', '藍色', '黃色']
|
| 152 |
color_name = color_map[i % len(color_map)]
|
| 153 |
output_text += f" {i+1}. {label} (信心度: {score:.2f}, 顏色: {color_name})\n"
|
| 154 |
-
|
| 155 |
if 'error' in debug_info:
|
| 156 |
output_text += f"\n錯誤: {debug_info['error']}"
|
| 157 |
|
|
@@ -159,7 +158,7 @@ def segment_and_display(image, text):
|
|
| 159 |
|
| 160 |
|
| 161 |
with gr.Blocks() as demo:
|
| 162 |
-
gr.Markdown("# Text-Guided Image Segmentation")
|
| 163 |
gr.Markdown("""
|
| 164 |
### 使用說明
|
| 165 |
1. 上傳一張圖片 (有提供預設圖片方便 demo)
|
|
@@ -170,7 +169,7 @@ with gr.Blocks() as demo:
|
|
| 170 |
with gr.Row():
|
| 171 |
with gr.Column(scale=1):
|
| 172 |
# 預設圖片 car.jpg
|
| 173 |
-
image_input = gr.Image(label="Input Image", value="car.jpg")
|
| 174 |
text_input = gr.Textbox(
|
| 175 |
label="Text Prompt",
|
| 176 |
placeholder="e.g. 'car. sky. road.'",
|
|
|
|
| 142 |
|
| 143 |
result_image, debug_info = segment_image_with_text(image, text)
|
| 144 |
|
| 145 |
+
output_text = f"檢測到 {debug_info.get('num_detections', 0)} 個物件\n---\n"
|
| 146 |
+
output_text += f"輸入文字: '{debug_info.get('original_prompt', text)}'\n---\n"
|
| 147 |
|
| 148 |
if 'labels' in debug_info and len(debug_info['labels']) > 0:
|
| 149 |
output_text += "檢測結果:\n"
|
|
|
|
| 151 |
color_map = ['紅色', '綠色', '藍色', '黃色']
|
| 152 |
color_name = color_map[i % len(color_map)]
|
| 153 |
output_text += f" {i+1}. {label} (信心度: {score:.2f}, 顏色: {color_name})\n"
|
|
|
|
| 154 |
if 'error' in debug_info:
|
| 155 |
output_text += f"\n錯誤: {debug_info['error']}"
|
| 156 |
|
|
|
|
| 158 |
|
| 159 |
|
| 160 |
with gr.Blocks() as demo:
|
| 161 |
+
gr.Markdown("# Text-Guided Image Segmentation Demo")
|
| 162 |
gr.Markdown("""
|
| 163 |
### 使用說明
|
| 164 |
1. 上傳一張圖片 (有提供預設圖片方便 demo)
|
|
|
|
| 169 |
with gr.Row():
|
| 170 |
with gr.Column(scale=1):
|
| 171 |
# 預設圖片 car.jpg
|
| 172 |
+
image_input = gr.Image(label="Input Image", value="sample_images/car.jpg")
|
| 173 |
text_input = gr.Textbox(
|
| 174 |
label="Text Prompt",
|
| 175 |
placeholder="e.g. 'car. sky. road.'",
|
car.jpg → sample_images/car.jpg
RENAMED
|
File without changes
|