app.py

from __future__ import annotations

from typing import Any, Dict

import numpy as np
import torch
from PIL import Image
import gradio as gr
from transformers import (
    AutoModelForZeroShotObjectDetection,
    AutoProcessor,
    SamModel,
    SamProcessor,
)

GROUNDING_MODEL_ID = "IDEA-Research/grounding-dino-base"
SAM_MODEL_ID = "facebook/sam-vit-base"

# 載入模型（使用 CPU）
device = "cpu"
grounding_processor = AutoProcessor.from_pretrained(GROUNDING_MODEL_ID)
grounding_model = AutoModelForZeroShotObjectDetection.from_pretrained(GROUNDING_MODEL_ID).to(device)
sam_processor = SamProcessor.from_pretrained(SAM_MODEL_ID)
sam_model = SamModel.from_pretrained(SAM_MODEL_ID).to(device)


def segment_image_with_text(
    image: Image.Image,
    text_prompt: str) -> tuple[np.ndarray, Dict[str, Any]]:
    """
    使用 Grounding DINO 檢測物件，然後使用 SAM 進行分割
    
    Args:
        image: PIL Image
        text_prompt: 文字提示
        
    Returns:
        tuple: (分割遮罩, 除錯資訊)
    """
    try:
        # 格式化文字提示：確保每個詞後面都有句號
        # Grounding DINO 期望的格式是 "object1. object2. object3."，沒有句號會有偵測不到的問題
        formatted_prompt = text_prompt.strip()
        if not formatted_prompt.endswith('.'):
            formatted_prompt += '.'
        
        # 步驟 1: 使用 Grounding DINO 檢測物件
        inputs = grounding_processor(images=image, text=formatted_prompt, return_tensors="pt").to(device)
        
        with torch.no_grad():
            outputs = grounding_model(**inputs)
        
        # 後處理檢測結果
        results = grounding_processor.post_process_grounded_object_detection(
            outputs,
            inputs.input_ids,
            threshold=0.15,  # 降低閾值以檢測更多物件
            target_sizes=[image.size[::-1]]
        )
        
        boxes = results[0]["boxes"].cpu().numpy()
        scores = results[0]["scores"].cpu().numpy()
        labels = results[0]["labels"]
        
        debug_info = {
            "num_detections": len(boxes),
            "scores": scores.tolist(),
            "labels": labels,
            "boxes": boxes.tolist(),
            "original_prompt": text_prompt
        }
        
        # 準備圖像陣列
        image_array = np.array(image)
        if image_array.shape[-1] == 4:  # 如果有 alpha 通道
            image_array = image_array[..., :3]
        
        if len(boxes) == 0:
            # 沒有檢測到物件，返回原圖
            return image_array, debug_info
        
        # 步驟 2: 為每個檢測框使用 SAM 進行分割
        overlay = image_array.copy()
        
        # 為每個檢測到的物件生成不同顏色的遮罩
        colors = [
            [255, 0, 0],    # 紅色
            [0, 255, 0],    # 綠色
            [0, 0, 255],    # 藍色
            [255, 255, 0],  # 黃色
        ]
        
        for idx, (box, label, score) in enumerate(zip(boxes, labels, scores)):
            # 為每個框單獨使用 SAM
            # box 格式應該是 [x_min, y_min, x_max, y_max]
            sam_inputs = sam_processor(
                image,
                input_boxes=[[box.tolist()]],
                return_tensors="pt"
            ).to(device)
            
            with torch.no_grad():
                sam_outputs = sam_model(**sam_inputs)
            
            # 取得分割遮罩
            masks = sam_processor.image_processor.post_process_masks(
                sam_outputs.pred_masks.cpu(),
                sam_inputs["original_sizes"].cpu(),
                sam_inputs["reshaped_input_sizes"].cpu()
            )
            
            # 取得第一個遮罩並轉換為 numpy
            mask = masks[0].squeeze().numpy()
            if mask.ndim == 3:
                mask = mask[0]  # 取第一個遮罩
            mask = mask > 0
            
            # 使用不同顏色標示不同物件
            color = colors[idx % len(colors)]
            mask_overlay = np.zeros_like(image_array)
            mask_overlay[mask] = color
            
            # 混合到結果圖像
            overlay[mask] = overlay[mask] * 0.5 + mask_overlay[mask] * 0.5
        
        return overlay.astype(np.uint8), debug_info
        
    except Exception as e:
        debug_info = {"error": str(e)}
        return np.zeros((480, 640, 3), dtype=np.uint8), debug_info


def segment_and_display(image, text):
    if image is None:
        return None, "請上傳圖片"
    if not text or text.strip() == "":
        return None, "請輸入文字提示"
    
    # 轉換為 PIL Image（如果需要）
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image)
    
    result_image, debug_info = segment_image_with_text(image, text)
    
    output_text = f"檢測到 {debug_info.get('num_detections', 0)} 個物件\n---\n"
    output_text += f"輸入文字: '{debug_info.get('original_prompt', text)}'\n---\n"
    
    if 'labels' in debug_info and len(debug_info['labels']) > 0:
        output_text += "檢測結果:\n"
        for i, (label, score) in enumerate(zip(debug_info['labels'], debug_info.get('scores', []))):
            color_map = ['紅色', '綠色', '藍色', '黃色']
            color_name = color_map[i % len(color_map)]
            output_text += f"  {i+1}. {label} (信心度: {score:.2f}, 顏色: {color_name})\n"
    if 'error' in debug_info:
        output_text += f"\n錯誤: {debug_info['error']}"
    
    return result_image, output_text


with gr.Blocks() as demo:
    gr.Markdown("# Text-Guided Image Segmentation Demo")
    gr.Markdown("""
    ### 使用說明
    1. 上傳一張圖片 (有提供預設圖片方便 demo)
    2. 輸入文字描述（例如：car、sky、road）
    3. 多個物件請用句號分隔（例如：car. sky. road.）
    """)

    with gr.Row():
        with gr.Column(scale=1):
            # 預設圖片 car.jpg
            image_input = gr.Image(label="Input Image", value="sample_images/car.jpg")
            text_input = gr.Textbox(
                label="Text Prompt", 
                placeholder="e.g. 'car. sky. road.'",
                lines=1
            )
            with gr.Row():
                segment_button = gr.Button("Segment", variant="primary")
                clear_button = gr.Button("Clear")
        with gr.Column(scale=2):
            output_mask = gr.Image(label="Segmentation Mask", type="numpy")
            summary = gr.Textbox(
                label="Summary", 
                interactive=False,
                lines=3,
                max_lines=20,
                show_copy_button=True
            )

    text_input.submit(segment_and_display, inputs=[image_input, text_input], outputs=[output_mask, summary])
    segment_button.click(segment_and_display, inputs=[image_input, text_input], outputs=[output_mask, summary])
    clear_button.click(lambda: (None, "", None, ""), inputs=None, outputs=[image_input, text_input, output_mask, summary])


demo.launch()