sam3-demo / app.py
danylgao
update
73441d8
#!/usr/bin/env python3
"""
SAM 3 目标检测与分割演示的 Hugging Face Spaces 版本。
适配 Hugging Face Spaces 部署环境:
1. 直接从 Hugging Face Hub 下载模型和资源
2. 支持 ZeroGPU 或 CPU 推理
3. 无需本地上传额外文件
支持功能:
1. 文本提示分割
2. 单框/多框提示分割
3. 正框/负框交互式标注(Multi Box 模式下可切换绘制正框或负框)
"""
import os
import torch
import numpy as np
import gradio as gr
from PIL import Image, ImageDraw, ImageFont
import matplotlib.pyplot as plt
import io
import random
from typing import List, Dict, Any, Tuple
# Hugging Face Hub 下载工具
from huggingface_hub import hf_hub_download, snapshot_download
# Hugging Face Spaces 环境检测
IS_HF_SPACES = os.environ.get("SPACE_ID") is not None
# 尝试导入 ImagePrompter,如果失败则使用标准 gr.Image
try:
from gradio_image_prompter import ImagePrompter
IMAGE_PROMPTER_AVAILABLE = True
except ImportError as e:
print(f"ImagePrompter 不可用: {e}")
IMAGE_PROMPTER_AVAILABLE = False
# 尝试导入 spaces 模块(用于 ZeroGPU)
try:
import spaces
SPACES_GPU_AVAILABLE = True
except ImportError:
SPACES_GPU_AVAILABLE = False
print("Hugging Face Spaces GPU 模块不可用,将使用标准推理")
# --- Hugging Face Hub 配置 ---
# SAM3 官方仓库:
# - HuggingFace: https://huggingface.co/facebook/sam3
# - GitHub: https://github.com/facebookresearch/sam3
SAM3_HF_REPO_ID = os.environ.get("SAM3_HF_REPO_ID", "facebook/sam3")
# 导入 sam3 库(通过 requirements.txt 从 GitHub 自动安装)
# requirements.txt 中配置: git+https://github.com/facebookresearch/sam3.git
SAM3_INSTALLED = False
sam3 = None
build_sam3_image_model = None
box_xywh_to_cxcywh = None
Sam3Processor = None
normalize_bbox = None
draw_box_on_image = None
plot_mask = None
plot_bbox = None
COLORS = [(1, 0, 0), (0, 1, 0), (0, 0, 1)] # 默认颜色
plot_results = None
try:
import sam3
from sam3 import build_sam3_image_model
SAM3_INSTALLED = True
print("✅ sam3 库已安装")
# 尝试导入其他组件
try:
from sam3.model.box_ops import box_xywh_to_cxcywh
except ImportError as e:
print(f"⚠️ box_ops 导入失败: {e}")
# 定义一个简单的替代函数
def box_xywh_to_cxcywh(boxes):
"""将 XYWH 格式转换为 CXCYWH 格式"""
x, y, w, h = boxes.unbind(-1)
cx = x + w / 2
cy = y + h / 2
return torch.stack([cx, cy, w, h], dim=-1)
try:
from sam3.model.sam3_image_processor import Sam3Processor
except ImportError as e:
print(f"⚠️ Sam3Processor 导入失败: {e}")
Sam3Processor = None
try:
from sam3.visualization_utils import normalize_bbox, draw_box_on_image, plot_mask, plot_bbox, COLORS, plot_results
except ImportError as e:
print(f"⚠️ visualization_utils 导入失败: {e}")
# 定义简单的替代函数
def normalize_bbox(boxes, width, height):
"""归一化边界框坐标"""
if isinstance(boxes, torch.Tensor):
normalized = boxes.clone()
normalized[..., 0] /= width
normalized[..., 1] /= height
normalized[..., 2] /= width
normalized[..., 3] /= height
return normalized
return boxes
def plot_mask(mask, color=(1, 0, 0), alpha=0.5):
"""绘制掩码"""
import matplotlib.pyplot as plt
h, w = mask.shape[-2:]
mask_image = mask.reshape(h, w, 1) * np.array(color).reshape(1, 1, -1)
plt.imshow(mask_image, alpha=alpha)
def plot_bbox(h, w, box, text="", box_format="XYXY", color=(1, 0, 0), relative_coords=False):
"""绘制边界框"""
import matplotlib.pyplot as plt
import matplotlib.patches as patches
if isinstance(box, torch.Tensor):
box = box.tolist()
x0, y0, x1, y1 = box
rect = patches.Rectangle((x0, y0), x1-x0, y1-y0, linewidth=2, edgecolor=color, facecolor='none')
plt.gca().add_patch(rect)
if text:
plt.text(x0, y0, text, color=color, fontsize=8)
COLORS = [(1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 0), (1, 0, 1), (0, 1, 1)]
plot_results = None
draw_box_on_image = None
except ImportError as e:
print(f"❌ sam3 库导入失败: {e}")
print("请确保 requirements.txt 中包含: git+https://github.com/facebookresearch/sam3.git")
# --- 0. ImagePrompter 数据解析函数 ---
def draw_boxes_with_labels(
image: Image.Image,
xyxy_boxes: List[List[float]],
box_labels: List[bool]
) -> Image.Image:
"""
在图像上绘制带颜色和标签的框。
Args:
image: 原始 PIL 图像
xyxy_boxes: 框坐标列表 [[x_min, y_min, x_max, y_max], ...]
box_labels: 框标签列表 [True/False, ...],True=正框(绿色),False=负框(红色)
Returns:
带有彩色框和标签的图像
"""
if image is None:
return None
# 复制图像以避免修改原图
img_draw = image.copy()
draw = ImageDraw.Draw(img_draw)
# 尝试加载字体,如果失败则使用默认字体
try:
font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 16)
except:
try:
font = ImageFont.truetype("Arial.ttf", 16)
except:
font = ImageFont.load_default()
for i, (box, label) in enumerate(zip(xyxy_boxes, box_labels)):
x_min, y_min, x_max, y_max = [int(coord) for coord in box]
# 正框用红色,负框用绿色
if label:
color = (255, 0, 0) # 红色 - 正框
label_text = f"Box {i}: True (正框)"
else:
color = (0, 255, 0) # 绿色 - 负框
label_text = f"Box {i}: False (负框)"
# 绘制矩形框,线宽为3
draw.rectangle([x_min, y_min, x_max, y_max], outline=color, width=3)
# 绘制标签背景
text_bbox = draw.textbbox((x_min, y_min - 20), label_text, font=font)
# 确保标签在图像内
text_y = max(0, y_min - 22)
if text_y == 0:
text_y = y_max + 2 # 如果顶部没空间,放到框下方
text_bbox = draw.textbbox((x_min, text_y), label_text, font=font)
draw.rectangle(text_bbox, fill=color)
draw.text((x_min, text_y), label_text, fill="white", font=font)
return img_draw
def process_imageprompter_data(
data: Any,
box_mode_history: List[Tuple[int, str]] = None,
verbose: bool = False
) -> Tuple[List[List[float]], List[bool]]:
"""
处理 ImagePrompter 数据,提取框坐标 (XYXY 格式) 和对应的标签(正/负框)。
ImagePrompter 返回格式:
{'image': <PIL Image>, 'points': [[x1, y1, label1, x2, y2, label2], ...]}
Args:
data: ImagePrompter 返回的数据字典
box_mode_history: 框模式切换历史列表,格式为 [(框索引, 模式), ...]
例如 [(0, "positive"), (2, "negative")] 表示第0个框开始是正框,第2个框开始是负框
如果为 None 或空,则所有框默认为正框
verbose: 是否输出详细调试日志
Returns:
tuple: (xyxy_boxes, box_labels)
- xyxy_boxes: 框坐标列表 [[x_min, y_min, x_max, y_max], ...]
- box_labels: 框标签列表 [True/False, ...],True=正框,False=负框
"""
if data is None or not isinstance(data, dict):
return [], []
xyxy_boxes = []
if verbose:
print(f"\n--- Shape Parsing Debug START ---")
print(f"Debug: Data keys = {list(data.keys())}")
print(f"Debug: Box mode history = {box_mode_history}")
# 从 'points' 键提取框(ImagePrompter 主要格式)
# 格式: [[x1, y1, label1, x2, y2, label2], ...]
if 'points' in data and data['points'] is not None:
points_list = data['points']
for i, points in enumerate(points_list):
if isinstance(points, (list, np.ndarray)) and len(points) >= 6:
try:
# ImagePrompter 格式: [x1, y1, label1, x2, y2, label2]
x1 = float(points[0])
y1 = float(points[1])
x2 = float(points[3])
y2 = float(points[4])
# 确保坐标顺序正确 (min, min, max, max)
x_min = min(x1, x2)
x_max = max(x1, x2)
y_min = min(y1, y2)
y_max = max(y1, y2)
box = [x_min, y_min, x_max, y_max]
xyxy_boxes.append(box)
except (ValueError, TypeError, IndexError):
pass # 跳过无效的点数据
# 根据 box_mode_history 生成标签列表
# 策略:根据历史记录中的切换点,确定每个框的模式
box_labels = []
current_mode = "positive" # 默认为正框模式
# 构建一个映射:框索引 -> 模式
mode_switch_points = {}
if box_mode_history:
for box_idx, mode in box_mode_history:
mode_switch_points[box_idx] = mode
for i in range(len(xyxy_boxes)):
# 检查是否在此索引处有模式切换
if i in mode_switch_points:
current_mode = mode_switch_points[i]
is_positive = (current_mode == "positive")
box_labels.append(is_positive)
if verbose:
print(f"Total boxes: {len(xyxy_boxes)} (正框: {sum(box_labels) if box_labels else 0}, 负框: {len(box_labels) - sum(box_labels) if box_labels else 0})")
print(f"--- Shape Parsing Debug END ---\n")
return xyxy_boxes, box_labels
# --- 1. 辅助函数 ---
def plot_boxes_to_image(
image_pil: Image,
tgt: Dict,
return_point: bool = False,
point_width: float = 1.0,
return_score=True,
) -> Image:
"""Plot bounding boxes and labels on an image."""
boxes = tgt["boxes"]
scores = tgt["scores"]
draw = ImageDraw.Draw(image_pil)
mask = Image.new("L", image_pil.size, 0)
mask_draw = ImageDraw.Draw(mask)
for box, score in zip(boxes, scores):
color = tuple(np.random.randint(0, 255, size=3).tolist())
x0, y0, x1, y1 = box
x0, y0, x1, y1 = int(x0), int(y0), int(x1), int(y1)
if return_point:
center_x = int((x0 + x1) / 2)
center_y = int((y0 + y1) / 2)
draw.ellipse(
(
center_x - point_width,
center_y - point_width,
center_x + point_width,
center_y + point_width,
),
fill=color,
width=point_width,
)
else:
draw.rectangle([x0, y0, x1, y1], outline=color, width=int(point_width))
if return_score:
text = f"{score:.2f}"
else:
text = f""
font = ImageFont.load_default()
if hasattr(font, "getbbox"):
bbox = draw.textbbox((x0, y0), text, font)
else:
w, h = draw.textsize(text, font)
bbox = (x0, y0, w + x0, y0 + h)
if not return_point:
draw.rectangle(bbox, fill=color)
draw.text((x0, y0), text, fill="white")
mask_draw.rectangle([x0, y0, x1, y1], fill=255, width=6)
return image_pil, mask
def parse_visual_prompt(points: List):
"""Parse visual prompt points to bounding boxes (XYXY format)"""
boxes = []
pos_points = []
neg_points = []
for point in points:
if point[2] == 2 and point[-1] == 3:
x1, y1, _, x2, y2, _ = point
boxes.append([x1, y1, x2, y2])
elif point[2] == 1 and point[-1] == 4:
x, y, _, _, _, _ = point
pos_points.append([x, y])
elif point[2] == 0 and point[-1] == 4:
x, y, _, _, _, _ = point
neg_points.append([x, y])
return boxes, pos_points, neg_points
# --- 2. 模型和处理器初始化 ---
# 从 Hugging Face Hub 下载所有必要资源
bpe_path = None
sam3_checkpoint = None
example_image_hf_path = None
def download_resources_from_hf():
"""从 Hugging Face Hub 下载模型和资源文件"""
global bpe_path, sam3_checkpoint, example_image_hf_path
if not SAM3_INSTALLED:
print("❌ sam3 库未安装,无法下载资源")
return False
try:
# 1. 下载 BPE 词汇表
bpe_path = hf_hub_download(
repo_id=SAM3_HF_REPO_ID,
filename="assets/bpe_simple_vocab_16e6.txt.gz",
cache_dir=os.environ.get("HF_HOME", None)
)
print(f"✅ BPE 词汇表: {bpe_path}")
except Exception as e:
print(f"⚠️ 无法下载 BPE 词汇表: {e}")
# 尝试从本地 sam3 模块获取
if sam3 is not None:
sam3_root = os.path.join(os.path.dirname(sam3.__file__), "..")
bpe_path = os.path.join(sam3_root, "assets", "bpe_simple_vocab_16e6.txt.gz")
if not os.path.exists(bpe_path):
bpe_path = None
try:
# 2. 下载模型检查点
# 优先使用环境变量指定的路径
env_checkpoint = os.environ.get("SAM3_CHECKPOINT_PATH")
if env_checkpoint and os.path.exists(env_checkpoint):
sam3_checkpoint = env_checkpoint
print(f"✅ 使用环境变量指定的模型: {sam3_checkpoint}")
else:
# 从 HF Hub 下载
sam3_checkpoint = hf_hub_download(
repo_id=SAM3_HF_REPO_ID,
filename="checkpoints/sam3.pt", # 或 "sam3.pt",根据实际仓库结构调整
cache_dir=os.environ.get("HF_HOME", None)
)
print(f"✅ 模型检查点: {sam3_checkpoint}")
except Exception as e:
print(f"⚠️ 无法下载模型检查点: {e}")
# 尝试其他文件名
try:
sam3_checkpoint = hf_hub_download(
repo_id=SAM3_HF_REPO_ID,
filename="sam3.pt",
cache_dir=os.environ.get("HF_HOME", None)
)
print(f"✅ 模型检查点(备选): {sam3_checkpoint}")
except:
sam3_checkpoint = None
try:
# 3. 下载示例图片
example_image_hf_path = hf_hub_download(
repo_id=SAM3_HF_REPO_ID,
filename="assets/images/test_image.jpg",
cache_dir=os.environ.get("HF_HOME", None)
)
print(f"✅ 示例图片: {example_image_hf_path}")
except Exception as e:
print(f"⚠️ 无法下载示例图片: {e}")
example_image_hf_path = None
return bpe_path is not None and sam3_checkpoint is not None
# 启动时下载资源
print(f"\n{'='*50}")
print(f"正在从 Hugging Face Hub 下载资源...")
print(f"仓库 ID: {SAM3_HF_REPO_ID}")
print(f"{'='*50}\n")
download_resources_from_hf()
# 设备配置
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {DEVICE}")
# 全局模型变量(延迟加载)
model = None
processor = None
autocast_ctx = None
def load_model():
"""延迟加载模型(支持 ZeroGPU)"""
global model, processor, autocast_ctx
if model is not None:
return True
if not SAM3_INSTALLED:
print("❌ sam3 库未安装")
return False
if sam3_checkpoint is None:
print("❌ 模型检查点路径未配置")
return False
if bpe_path is None:
print("❌ BPE 词汇表路径未配置")
return False
try:
if DEVICE == "cuda":
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
autocast_ctx = torch.autocast("cuda", dtype=torch.bfloat16)
autocast_ctx.__enter__()
model = build_sam3_image_model(bpe_path=bpe_path, checkpoint_path=sam3_checkpoint).to(DEVICE)
else:
autocast_ctx = None
model = build_sam3_image_model(bpe_path=bpe_path, checkpoint_path=sam3_checkpoint).to(DEVICE)
processor = Sam3Processor(model, confidence_threshold=0.5)
print("✅ 模型加载成功")
return True
except Exception as e:
print(f"❌ 模型加载失败: {e}")
import traceback
traceback.print_exc()
model = None
processor = None
return False
# 非 ZeroGPU 环境下预加载模型
if not SPACES_GPU_AVAILABLE:
load_model()
# --- 3. 可视化辅助函数 ---
def plot_to_pil(fig):
"""将 Matplotlib 图形转换为 PIL Image。"""
buf = io.BytesIO()
fig.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
buf.seek(0)
plt.close(fig)
return Image.open(buf).convert("RGB")
def get_result_figure(
img: Image.Image,
results: dict,
return_point: bool = False,
point_width: float = 3.0,
return_score: bool = True
) -> Tuple[plt.Figure, int]:
"""封装原始 plot_results 逻辑,支持显示中心点和置信度控制。
Args:
img: 输入图像
results: 推理结果字典
return_point: 是否显示中心点而不是边框
point_width: 中心点或边框线宽
return_score: 是否显示置信度分数
"""
fig = plt.figure(figsize=(12, 8))
plt.imshow(img)
plt.axis("off")
nb_objects = len(results.get("scores", []))
print(f"found {nb_objects} object(s)")
for i in range(nb_objects):
color = COLORS[i % len(COLORS)]
if "masks" in results and i < len(results["masks"]):
mask_data = results["masks"][i]
if mask_data.ndim == 3:
mask_data = mask_data.squeeze(0)
plot_mask(mask_data.cpu(), color=color)
if "boxes" in results and i < len(results["boxes"]):
w, h = img.size
box = results["boxes"][i].cpu().tolist()
prob = results["scores"][i].item()
# 根据参数决定显示方式
if return_point:
# 显示中心点
x0, y0, x1, y1 = box
center_x = (x0 + x1) / 2
center_y = (y0 + y1) / 2
circle = plt.Circle(
(center_x, center_y),
point_width * 2,
color=color,
fill=True
)
plt.gca().add_patch(circle)
# 如果显示置信度,在中心点旁边显示
if return_score:
plt.text(
center_x + point_width * 3,
center_y,
f"{prob:.2f}",
color=color,
fontsize=10,
fontweight='bold'
)
else:
# 显示边框
text = f"(id={i}, {prob:.2f})" if return_score else f"(id={i})"
plot_bbox(
h,
w,
results["boxes"][i].cpu(),
text=text,
box_format="XYXY",
color=color,
relative_coords=False,
)
return fig, nb_objects
# --- 4. Gradio 推理函数 ---
def sam3_segmentation_core(
unified_image_input: Any,
prompt_text: str,
box_type: str,
box_mode_history: List[Tuple[int, str]],
return_point: bool = False,
point_width: float = 3.0,
return_score: bool = True
):
"""核心分割函数"""
global model, processor
# 确保模型已加载
if not SAM3_INSTALLED:
return None, "❌ sam3 库未安装,请检查 requirements.txt 或 HF 仓库配置。", box_mode_history
if model is None or processor is None:
if not load_model():
return None, "❌ 模型未加载,请检查模型配置。可能需要设置 SAM3_HF_REPO_ID 环境变量。", box_mode_history
# 1. 从统一输入中提取图像和框数据
image = None
visual_prompter_data = None
if IMAGE_PROMPTER_AVAILABLE and isinstance(unified_image_input, dict):
image = unified_image_input.get('image')
visual_prompter_data = unified_image_input
else:
image = unified_image_input
visual_prompter_data = None
if image is None:
return None, "请上传图像。", box_mode_history
img0 = image.copy()
width, height = img0.size
# 2. 图像预处理
try:
inference_state = processor.set_image(img0)
except Exception as e:
return None, f"图像处理失败: {e}", box_mode_history
# 3. 清除并设置提示
processor.reset_all_prompts(inference_state)
found_objects = 0
xyxy_boxes = []
# --- 文本提示 ---
if box_type == "Text":
if not prompt_text:
return None, "文本模式下,请提供文本提示。", box_mode_history
inference_state = processor.set_text_prompt(
state=inference_state,
prompt=prompt_text
)
caption_base = "文本提示分割"
# --- 框提示 (Single/Multi Box) ---
elif box_type in ["Single Box", "Multi Box"]:
if not IMAGE_PROMPTER_AVAILABLE:
return None, "当前环境不支持 ImagePrompter,Box 模式无法运行。", box_mode_history
if visual_prompter_data:
# 调用解析函数,传入框模式历史(推理时启用详细日志)
xyxy_boxes, box_labels = process_imageprompter_data(visual_prompter_data, box_mode_history, verbose=True)
print(f"Boxes: {xyxy_boxes}")
print(f"Labels: {box_labels}")
if not xyxy_boxes:
return None, f"请在图像上绘制至少一个矩形框作为提示(当前模式: {box_type})。", box_mode_history
# 针对 Single Box 模式,只取第一个框
if box_type == "Single Box" and len(xyxy_boxes) > 1:
xyxy_boxes = [xyxy_boxes[0]]
box_labels = [box_labels[0]] if box_labels else [True]
box_inputs = []
for i, (x_min, y_min, x_max, y_max) in enumerate(xyxy_boxes):
x = x_min
y = y_min
w = x_max - x_min
h = y_max - y_min
box_inputs.append([x, y, w, h])
# 4. 转换并添加提示
try:
box_input_xywh = torch.tensor(box_inputs, dtype=torch.float32).view(-1, 4).to(DEVICE)
box_input_cxcywh = box_xywh_to_cxcywh(box_input_xywh)
norm_boxes_cxcywh = normalize_bbox(box_input_cxcywh, width, height).tolist()
for i in range(len(box_inputs)):
norm_box = norm_boxes_cxcywh[i]
label = box_labels[i] if i < len(box_labels) else True
label_str = "正框" if label else "负框"
print(f"Adding box {i}: {norm_box}, label={label} ({label_str})")
# 注意参数顺序: add_geometric_prompt(box, label, state)
inference_state = processor.add_geometric_prompt(
state=inference_state, box=norm_box, label=label
)
except Exception as e:
print(f"Error during box conversion/prompt setting: {e}")
return None, f"框提示处理失败: {e}", box_mode_history
num_positive = sum(box_labels) if box_labels else len(xyxy_boxes)
num_negative = len(xyxy_boxes) - num_positive
caption_base = f"使用 {len(xyxy_boxes)} 个提示框分割(正框: {num_positive}, 负框: {num_negative})"
else:
return None, "请选择有效的提示类型 (Text, Single Box, 或 Multi Box)。", box_mode_history
# 5. 运行推理和可视化
fig, found_objects = get_result_figure(
img0.copy(),
inference_state,
return_point=return_point,
point_width=point_width,
return_score=return_score
)
result_image = plot_to_pil(fig)
return result_image, f"{caption_base}。找到 {found_objects} 个对象。", box_mode_history
# 根据是否支持 ZeroGPU 选择推理函数
if SPACES_GPU_AVAILABLE:
@spaces.GPU
def sam3_segmentation(
unified_image_input: Any,
prompt_text: str,
box_type: str,
box_mode_history: List[Tuple[int, str]],
return_point: bool = False,
point_width: float = 3.0,
return_score: bool = True
):
"""ZeroGPU 版本的推理函数"""
return sam3_segmentation_core(
unified_image_input, prompt_text, box_type,
box_mode_history, return_point, point_width, return_score
)
else:
def sam3_segmentation(
unified_image_input: Any,
prompt_text: str,
box_type: str,
box_mode_history: List[Tuple[int, str]],
return_point: bool = False,
point_width: float = 3.0,
return_score: bool = True
):
"""标准版本的推理函数"""
return sam3_segmentation_core(
unified_image_input, prompt_text, box_type,
box_mode_history, return_point, point_width, return_score
)
# --- 5. 框模式切换处理函数 ---
def on_box_mode_change(
new_mode: str,
unified_image_input: Any,
current_history: List[Tuple[int, str]]
) -> Tuple[List[Tuple[int, str]], str, Image.Image]:
"""
当用户切换框模式时,记录当前框数量和新模式,并更新预览。
Args:
new_mode: 新选择的模式 ("正框 (Positive)" 或 "负框 (Negative)")
unified_image_input: 当前 ImagePrompter 的数据
current_history: 当前的模式切换历史
Returns:
tuple: (更新后的历史, 状态信息文本, 预览图像)
"""
if current_history is None:
current_history = []
# 获取当前已绘制的框数量
current_box_count = 0
if unified_image_input and isinstance(unified_image_input, dict):
points = unified_image_input.get('points', [])
if points:
current_box_count = len(points)
# 转换模式名称
mode_internal = "positive" if "Positive" in new_mode or "正框" in new_mode else "negative"
# 添加新的切换点
# 记录:从第 current_box_count 个框开始,使用新模式
new_history = current_history.copy()
new_history.append((current_box_count, mode_internal))
# 生成状态信息
mode_display = "正框" if mode_internal == "positive" else "负框"
status = f"✅ 已切换到 {mode_display} 模式。从第 {current_box_count + 1} 个框开始将被标记为{mode_display}。"
print(f"Box mode changed: {new_mode} -> {mode_internal}, at box index {current_box_count}")
print(f"Updated history: {new_history}")
# 生成预览图像(使用 verbose=False 避免频繁日志输出)
preview_image = None
if unified_image_input and isinstance(unified_image_input, dict):
image = unified_image_input.get('image')
if image is not None:
xyxy_boxes, box_labels = process_imageprompter_data(unified_image_input, new_history, verbose=False)
if xyxy_boxes:
preview_image = draw_boxes_with_labels(image, xyxy_boxes, box_labels)
else:
preview_image = image
return new_history, status, preview_image
def reset_box_mode_history(
unified_image_input: Any
) -> Tuple[List[Tuple[int, str]], str, Image.Image]:
"""重置框模式历史并更新预览"""
new_history = [(0, "positive")]
status = "已重置,所有框将默认为正框。"
# 生成预览图像(使用 verbose=False 避免频繁日志输出)
preview_image = None
if unified_image_input and isinstance(unified_image_input, dict):
image = unified_image_input.get('image')
if image is not None:
xyxy_boxes, box_labels = process_imageprompter_data(unified_image_input, new_history, verbose=False)
if xyxy_boxes:
preview_image = draw_boxes_with_labels(image, xyxy_boxes, box_labels)
else:
preview_image = image
return new_history, status, preview_image
def get_current_box_status(
unified_image_input: Any,
box_mode_history: List[Tuple[int, str]]
) -> str:
"""获取当前框的状态信息"""
if not unified_image_input or not isinstance(unified_image_input, dict):
return "尚未绘制框"
points = unified_image_input.get('points', [])
if not points:
return "尚未绘制框"
num_boxes = len(points)
# 计算正负框数量
if not box_mode_history:
return f"已绘制 {num_boxes} 个框(全部为正框)"
# 根据历史计算每个框的标签
mode_switch_points = {}
for box_idx, mode in box_mode_history:
mode_switch_points[box_idx] = mode
current_mode = "positive"
positive_count = 0
negative_count = 0
for i in range(num_boxes):
if i in mode_switch_points:
current_mode = mode_switch_points[i]
if current_mode == "positive":
positive_count += 1
else:
negative_count += 1
return f"已绘制 {num_boxes} 个框(正框: {positive_count}, 负框: {negative_count})"
def update_box_preview(
unified_image_input: Any,
box_mode_history: List[Tuple[int, str]]
) -> Tuple[Image.Image, str, str]:
"""
更新框预览图像,显示带颜色和标签的框。
Args:
unified_image_input: ImagePrompter 的数据
box_mode_history: 框模式历史
Returns:
tuple: (预览图像, 状态文本, 框提示参数文本)
"""
# 获取状态文本
status_text = get_current_box_status(unified_image_input, box_mode_history)
# 检查输入有效性
if not unified_image_input or not isinstance(unified_image_input, dict):
return None, status_text, "None"
image = unified_image_input.get('image')
if image is None:
return None, status_text, "None"
# 解析框数据(使用 verbose=False 避免频繁日志输出)
xyxy_boxes, box_labels = process_imageprompter_data(unified_image_input, box_mode_history, verbose=False)
if not xyxy_boxes:
return image, status_text, "None"
# 绘制带颜色和标签的框
preview_image = draw_boxes_with_labels(image, xyxy_boxes, box_labels)
# 生成框提示参数文本
boxes_int = [[int(coord) for coord in box] for box in xyxy_boxes]
if len(xyxy_boxes) == 1:
prompt_info_text = f"Box: {boxes_int[0]}\nLabel: {box_labels[0]}"
else:
prompt_info_text = f"Boxes: {boxes_int}\nLabels: {box_labels}"
return preview_image, status_text, prompt_info_text
# --- 6. Gradio 接口定义 ---
# 示例图片加载(优先使用从 HF Hub 下载的图片)
example_image_path = None
example_image = None
# 优先级: HF Hub 下载 > 本地脚本目录 > sam3 模块目录 > 占位图
if example_image_hf_path and os.path.exists(example_image_hf_path):
example_image_path = example_image_hf_path
example_image = Image.open(example_image_hf_path)
print(f"✅ 使用 HF Hub 下载的示例图片: {example_image_path}")
else:
# 备选: 本地脚本目录
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
sam3_asset_path = os.path.join(SCRIPT_DIR, "assets", "images", "test_image.jpg")
if os.path.exists(sam3_asset_path):
example_image_path = os.path.abspath(sam3_asset_path)
example_image = Image.open(sam3_asset_path)
print(f"✅ 使用本地示例图片: {example_image_path}")
elif sam3 is not None:
# 备选: sam3 模块目录
sam3_root = os.path.join(os.path.dirname(sam3.__file__), "..")
sam3_asset_path = os.path.join(sam3_root, "assets", "images", "test_image.jpg")
if os.path.exists(sam3_asset_path):
example_image_path = os.path.abspath(sam3_asset_path)
example_image = Image.open(sam3_asset_path)
print(f"✅ 使用 sam3 模块示例图片: {example_image_path}")
if example_image is None:
print(f"⚠️ 示例图片未找到,使用占位图")
example_image_path = None
example_image = Image.new('RGB', (512, 512), color='lightgray')
# 示例数据:ImagePrompter 需要使用 "image" 和 "points" 格式
# points 格式: [[x1, y1, label1, x2, y2, label2], ...]
# 其中 label1=2, label2=3 表示框的起点和终点
# 示例框坐标和标签说明
# Single Box: Box=[487.0, 302.0, 591.0, 641.0], Label=True (正框)
# Multi Box: Boxes=[[487.0, 302.0, 591.0, 641.0], [341.0, 275.0, 495.0, 662.0]], Labels=[True, False]
# 用于示例展示的提示信息
example_prompts_info = {
"Text": "None",
"Single Box": "Box: [487, 302, 591, 641]\nLabel: True",
"Multi Box": "Boxes: [[487, 302, 591, 641], [341, 275, 495, 662]]\nLabels: [False, True]"
}
if IMAGE_PROMPTER_AVAILABLE:
# 注意:使用文件路径而非 PIL Image 对象,以便 Gradio Examples 正确显示缩略图
if example_image_path:
example_data_corrected = [
[{"image": example_image_path, "points": []}, "Text", example_prompts_info["Text"], "shoe"],
[{"image": example_image_path, "points": [[487.0, 302.0, 2, 591.0, 641.0, 3]]}, "Single Box", example_prompts_info["Single Box"], ""],
[{"image": example_image_path, "points": [[487.0, 302.0, 2, 591.0, 641.0, 3], [341.0, 275.0, 2, 495.0, 662.0, 3]]}, "Multi Box", example_prompts_info["Multi Box"], ""],
]
else:
example_data_corrected = [
[{"image": example_image, "points": []}, "Text", example_prompts_info["Text"], "shoe"],
[{"image": example_image, "points": [[487.0, 302.0, 2, 591.0, 641.0, 3]]}, "Single Box", example_prompts_info["Single Box"], ""],
[{"image": example_image, "points": [[487.0, 302.0, 2, 591.0, 641.0, 3], [341.0, 275.0, 2, 495.0, 662.0, 3]]}, "Multi Box", example_prompts_info["Multi Box"], ""],
]
# 设定 Multi Box 示例的默认历史:第0个框为负框,第1个框为正框
example_multi_box_history = [(0, "negative"), (1, "positive")]
else:
# 非 ImagePrompter 模式,使用路径或 PIL Image
if example_image_path:
example_data_corrected = [
[example_image_path, "Text", example_prompts_info["Text"], "shoe"],
[example_image_path, "Single Box", example_prompts_info["Single Box"], ""],
[example_image_path, "Multi Box", example_prompts_info["Multi Box"], ""],
]
else:
example_data_corrected = [
[example_image, "Text", example_prompts_info["Text"], "shoe"],
[example_image, "Single Box", example_prompts_info["Single Box"], ""],
[example_image, "Multi Box", example_prompts_info["Multi Box"], ""],
]
example_multi_box_history = [(0, "positive")]
def on_example_select(
unified_image_input: Any,
prompt_type: str
) -> Tuple[List[Tuple[int, str]], Image.Image, str]:
"""
当用户选择示例时,自动更新框模式历史和预览。
Args:
unified_image_input: ImagePrompter 的数据
prompt_type: 提示类型 (Text, Single Box, Multi Box)
Returns:
tuple: (框模式历史, 预览图像, 状态文本)
"""
# 根据提示类型设置框模式历史
if prompt_type == "Multi Box":
# Multi Box 示例: 第0个框为正框,第1个框为负框
box_history = [(0, "positive"), (1, "negative")]
elif prompt_type == "Single Box":
# Single Box 示例: 只有一个正框
box_history = [(0, "positive")]
else:
# Text 模式: 默认历史
box_history = [(0, "positive")]
# 生成预览图像
preview_image = None
status_text = "尚未绘制框"
if unified_image_input and isinstance(unified_image_input, dict):
image = unified_image_input.get('image')
if image is not None:
xyxy_boxes, box_labels = process_imageprompter_data(unified_image_input, box_history, verbose=False)
if xyxy_boxes:
preview_image = draw_boxes_with_labels(image, xyxy_boxes, box_labels)
num_positive = sum(box_labels)
num_negative = len(box_labels) - num_positive
status_text = f"已绘制 {len(xyxy_boxes)} 个框(正框: {num_positive}, 负框: {num_negative})"
else:
preview_image = image
return box_history, preview_image, status_text
# 构建 Gradio 界面
# 注意:为保证与各版本 Gradio 的兼容性,使用最简配置
with gr.Blocks() as demo:
# 状态变量:存储框模式切换历史
box_mode_history_state = gr.State([(0, "positive")]) # 默认从第0个框开始为正框模式
gr.Markdown(
"""
# 🎯 SAM 3 Demo
**Segment Anything Model 3 - 目标检测与分割**
> 🚀 Powered by Hugging Face Spaces
"""
)
with gr.Row():
with gr.Column(scale=1):
# 使用说明放在左侧第一行
with gr.Accordion("📋 使用说明", open=False):
gr.Markdown("""
**使用方法:**
📝 **Text 模式**
1. 选择 "Text" 模式
2. 上传图像
3. 输入文本提示词(如 "shoe", "person")
4. 点击"运行 SAM 3 分割"
⬜ **Single Box 模式**
1. 选择 "Single Box" 模式
2. 上传图像
3. 在图像上绘制一个矩形框
4. 点击"运行 SAM 3 分割"
🔲 **Multi Box 模式(支持正/负框)**
1. 选择 "Multi Box" 模式
2. 上传图像
3. **默认为正框模式**,绘制的框将包含目标
4. 如需绘制负框(排除区域):
- 先绘制正框
- 点击切换到 "负框 (Negative)" 模式
- 继续绘制负框
5. 点击「🔄 刷新预览」按钮查看框标签预览
6. 点击"运行 SAM 3 分割"
💡 **正框 vs 负框**
- **正框(红色)**: 告诉模型"包含这个区域的目标"
- **负框(绿色)**: 告诉模型"排除这个区域",用于去除误检
- **注意**: 负框需要配合正框使用才能生效
⚙️ **显示选项**
- **显示中心点**: 用圆点代替边框显示检测结果中心位置
- **显示置信度**: 在结果中显示模型的置信度分数
- **线条/点宽度**: 调整边框线宽或中心点大小
""")
if IMAGE_PROMPTER_AVAILABLE:
unified_image_input = ImagePrompter(
label="🖼️ 示例图像",
type="pil"
)
else:
unified_image_input = gr.Image(
label="🖼️ 示例图像",
type="pil"
)
prompt_type = gr.Radio(
["Text", "Single Box", "Multi Box"],
label="提示类型",
value="Text"
)
text_prompt_input = gr.Textbox(
label="文本提示参数",
value="shoe",
visible=True
)
# 添加一个用于显示框提示信息的文本框(仅用于示例展示)
example_prompt_info_display = gr.Textbox(
label="框提示参数",
value="",
interactive=False,
lines=2,
visible=True
)
# 框模式选择器(仅 Multi Box 模式显示)
with gr.Group(visible=False) as box_mode_group:
gr.Markdown("### 🎯 框模式设置")
box_mode_selector = gr.Radio(
["正框 (Positive)", "负框 (Negative)"],
label="当前绘制模式",
value="正框 (Positive)",
info="正框=包含目标,负框=排除区域"
)
with gr.Row():
reset_history_btn = gr.Button("🔄 重置框标签", size="sm")
# 框预览区域(Single Box 和 Multi Box 模式都显示)
with gr.Group(visible=False) as box_preview_group:
gr.Markdown("### 📦 框预览(红色=正框 True,绿色=负框 False)")
box_status_text = gr.Textbox(
label="框状态",
value="尚未绘制框",
interactive=False
)
refresh_preview_btn = gr.Button("🔄 刷新预览", size="sm", variant="secondary")
gr.Markdown("*绘制框后点击「刷新预览」按钮查看标注效果*")
box_preview_image = gr.Image(
label="框标签预览",
type="pil",
interactive=False
)
# 显示选项(调整到左侧)
gr.Markdown("### ⚙️ 显示选项")
with gr.Row():
return_point = gr.Checkbox(label="显示中心点", value=False)
return_score = gr.Checkbox(label="显示置信度", value=True)
point_width = gr.Slider(
label="线条/点宽度",
value=3.0,
minimum=0.0,
maximum=20.0,
step=0.1,
)
run_button = gr.Button("Run SAM3", variant="primary")
with gr.Column(scale=2):
output_image = gr.Image(label="分割结果", type="pil")
result_info = gr.Textbox(label="结果信息", lines=2)
def run_example(img, ptype, prompt_info, text):
"""运行示例时使用正确的框模式历史"""
if ptype == "Multi Box":
# Multi Box 示例: 第0个框为正框,第1个框为负框
history = [(0, "negative"), (1, "positive")]
else:
history = [(0, "positive")]
result_img, result_text, _ = sam3_segmentation(img, text, ptype, history, False, 3.0, True)
return result_img, result_text
gr.Examples(
examples=example_data_corrected,
inputs=[unified_image_input, prompt_type, example_prompt_info_display, text_prompt_input],
outputs=[output_image, result_info],
fn=run_example,
cache_examples=False,
label="示例"
)
# 事件绑定
run_button.click(
fn=sam3_segmentation,
inputs=[unified_image_input, text_prompt_input, prompt_type, box_mode_history_state, return_point, point_width, return_score],
outputs=[output_image, result_info, box_mode_history_state]
)
# 框模式切换事件(同时更新预览)
box_mode_selector.change(
fn=on_box_mode_change,
inputs=[box_mode_selector, unified_image_input, box_mode_history_state],
outputs=[box_mode_history_state, box_status_text, box_preview_image]
)
# 重置框历史(同时更新预览)
reset_history_btn.click(
fn=reset_box_mode_history,
inputs=[unified_image_input],
outputs=[box_mode_history_state, box_status_text, box_preview_image]
)
# 手动刷新预览按钮(避免 change 事件导致的持续更新问题)
# 同时更新框提示参数
refresh_preview_btn.click(
fn=update_box_preview,
inputs=[unified_image_input, box_mode_history_state],
outputs=[box_preview_image, box_status_text, example_prompt_info_display]
)
def update_inputs(p_type):
is_text = p_type == "Text"
is_multi_box = p_type == "Multi Box"
is_box_mode = p_type in ["Single Box", "Multi Box"] # Single Box 和 Multi Box 都显示预览
return (
gr.update(visible=is_text), # text_prompt_input
gr.update(visible=is_multi_box), # box_mode_group(正/负框切换,仅 Multi Box)
gr.update(visible=is_box_mode) # box_preview_group(框预览,Single/Multi Box 都显示)
)
def update_inputs_and_preview(p_type, img_input):
"""
更新输入组件可见性,并在示例加载时自动更新预览。
"""
is_text = p_type == "Text"
is_multi_box = p_type == "Multi Box"
is_box_mode = p_type in ["Single Box", "Multi Box"] # Single Box 和 Multi Box 都显示预览
# 根据提示类型设置框模式历史
if p_type == "Multi Box":
box_history = [(0, "negative"), (1, "positive")]
elif p_type == "Single Box":
box_history = [(0, "positive")]
else:
box_history = [(0, "positive")]
# 生成预览图像
preview_image = None
status_text = "尚未绘制框"
if img_input and isinstance(img_input, dict):
image = img_input.get('image')
if image is not None:
xyxy_boxes, box_labels = process_imageprompter_data(img_input, box_history, verbose=False)
if xyxy_boxes:
preview_image = draw_boxes_with_labels(image, xyxy_boxes, box_labels)
num_positive = sum(box_labels)
num_negative = len(box_labels) - num_positive
status_text = f"已绘制 {len(xyxy_boxes)} 个框(正框: {num_positive}, 负框: {num_negative})"
else:
preview_image = image
return (
gr.update(visible=is_text), # text_prompt_input
gr.update(visible=is_multi_box), # box_mode_group(正/负框切换,仅 Multi Box)
gr.update(visible=is_box_mode), # box_preview_group(框预览,Single/Multi Box 都显示)
box_history, # box_mode_history_state
preview_image, # box_preview_image
status_text # box_status_text
)
prompt_type.change(
fn=update_inputs_and_preview,
inputs=[prompt_type, unified_image_input],
outputs=[text_prompt_input, box_mode_group, box_preview_group, box_mode_history_state, box_preview_image, box_status_text]
)
# 注意:已移除 unified_image_input.change 事件,避免动态刷新问题
# 用户绘制框后需要手动点击「刷新预览」按钮查看预览
demo.load(
fn=update_inputs,
inputs=[prompt_type],
outputs=[text_prompt_input, box_mode_group, box_preview_group]
)
# Hugging Face Spaces 启动配置
if __name__ == "__main__":
# Hugging Face Spaces 会自动处理端口、地址和共享设置
# 不要指定 server_name 和 server_port,让平台自动配置
demo.launch(
show_error=True # 显示详细错误信息,方便调试
)