欢迎关注我的CSDN：https://spike.blog.csdn.net/
本文地址：https://spike.blog.csdn.net/article/details/143388189

免责声明：本文来源于个人知识与公开资料，仅用于学术交流，欢迎讨论，不支持转载。

---

Grounded SAM2 集成多个先进模型的视觉 AI 框架，融合 GroundingDINO、Florence-2 和 SAM2 等模型，实现开放域目标检测、分割和跟踪等多项视觉任务的突破性进展，通过自然语言描述来定位图像中的目标，生成精细的目标分割掩码，在视频序列中持续跟踪目标，保持 ID 的一致性。

Paper: Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks，SAM 版本由 1.0 升级至 2.0

1. 环境配置

GitHub: Grounded-SAM-2

git clone https://github.com/IDEA-Research/Grounded-SAM-2
cd Grounded-SAM-2

准备 SAM 2.1 模型，格式是 pt 的，GroundingDINO 模型，格式是 pth 的，即：

wget https://huggingface.co/facebook/sam2.1-hiera-large/resolve/main/sam2.1_hiera_large.pt?download=true -O sam2.1_hiera_large.pt
wget https://huggingface.co/ShilongLiu/GroundingDINO/resolve/main/groundingdino_swint_ogc.pth

最新模型位置：

cd checkpoints
ln -s [your path]/llm/workspace_comfyui/ComfyUI/models/sam2/sam2_hiera_large.pt sam2_hiera_large.pt

cd gdino_checkpoints
ln -s [your path]/llm/workspace_comfyui/ComfyUI/models/grounding-dino/groundingdino_swinb_cogcoor.pth groundingdino_swinb_cogcoor.pth
ln -s [your path]/llm/workspace_comfyui/ComfyUI/models/grounding-dino/groundingdino_swint_ogc.pth groundingdino_swint_ogc.pth

激活环境：

conda activate sam2

测试 PyTorch：

import torch
print(torch.__version__)  # 2.5.0+cu124
print(torch.cuda.is_available())  # True
exit()
echo $CUDA_HOME

安装 Grounding DINO：

pip install --no-build-isolation -e grounding_dino
pip show groundingdino

安装 SAM2：

pip install --no-build-isolation -e .
pip install --no-build-isolation -e ".[notebooks]"  # 适配 Jupyter
pip show SAM-2

配置参数：视觉分割开源算法 SAM2(Segment Anything 2) 配置与推理

依赖文件：

cd grounding_dino/
pip install -r requirements.txt --verbose

2. 测试图像

测试脚本：grounded_sam2_local_demo.py

导入相关的依赖包：

import os
import cv2
import json
import torch
import numpy as np
import supervision as sv
import pycocotools.mask as mask_util
from pathlib import Path
from torchvision.ops import box_convert
from sam2.build_sam import build_sam2
from sam2.sam2_image_predictor import SAM2ImagePredictor
from grounding_dino.groundingdino.util.inference import load_model, load_image, predict

from PIL import Image
import matplotlib.pyplot as plt

配置数据，以及依赖环境，其中包括：

• 输入文本提示，例如 袜子(socks) 和 吉他(guitar)
• 输入图像
• SAM2 模型 v2.1 版本，以及配置
• GroundingDINO (DETR with Improved deNoising anchOr boxes, 改进的去噪锚框的DETR) 模型，以及配置
• Box 阈值、文本阈值
• 输出文件夹与Json

即：

TEXT_PROMPT = "socks. guitar."
#IMG_PATH = "notebooks/images/truck.jpg"
IMG_PATH = "[your path]/llm/vision_test_data/image2.png"

image = Image.open(IMG_PATH)
plt.figure(figsize=(9, 6))
plt.title(f"annotated_frame")
plt.imshow(image)

SAM2_CHECKPOINT = "./checkpoints/sam2.1_hiera_large.pt"
SAM2_MODEL_CONFIG = "configs/sam2.1/sam2.1_hiera_l.yaml"
GROUNDING_DINO_CONFIG = "grounding_dino/groundingdino/config/GroundingDINO_SwinT_OGC.py"
GROUNDING_DINO_CHECKPOINT = "gdino_checkpoints/groundingdino_swint_ogc.pth"
BOX_THRESHOLD = 0.35
TEXT_THRESHOLD = 0.25
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
OUTPUT_DIR = Path("outputs/grounded_sam2_local_demo")
DUMP_JSON_RESULTS = True

# create output directory
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

加载 SAM2 模型，获得 sam2_predictor，即：

# build SAM2 image predictor
sam2_checkpoint = SAM2_CHECKPOINT
model_cfg = SAM2_MODEL_CONFIG
sam2_model = build_sam2(model_cfg, sam2_checkpoint, device=DEVICE)
sam2_predictor = SAM2ImagePredictor(sam2_model)

加载 GroundingDINO 模型，获得 grounding_model，即：

# build grounding dino model
grounding_model = load_model(
    model_config_path=GROUNDING_DINO_CONFIG, 
    model_checkpoint_path=GROUNDING_DINO_CHECKPOINT,
    device=DEVICE
)

SAM2 加载图像数据，即：

text = TEXT_PROMPT
img_path = IMG_PATH

# image(原图), image_transformed(正则化图像)
image_source, image = load_image(img_path)
sam2_predictor.set_image(image_source)

GroudingDINO 预测 Bounding Box，输入模型、图像、文本、Box和Text阈值，即：

• load_image() 和 predict() 都来自于 GroundingDINO，数据和模型匹配。

boxes, confidences, labels = predict(
    model=grounding_model,
    image=image,
    caption=text,
    box_threshold=BOX_THRESHOLD,
    text_threshold=TEXT_THRESHOLD,
)

适配不同 Box 的格式：

h, w, _ = image_source.shape
boxes = boxes * torch.Tensor([w, h, w, h])
input_boxes = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()

SAM2 依赖的 PyTorch 配置：

# FIXME: figure how does this influence the G-DINO model
torch.autocast(device_type="cuda", dtype=torch.bfloat16).__enter__()

if torch.cuda.get_device_properties(0).major >= 8:
    # turn on tfloat32 for Ampere GPUs (https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices)
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.allow_tf32 = True

SAM2 预测图像：

masks, scores, logits = sam2_predictor.predict(
    point_coords=None,
    point_labels=None,
    box=input_boxes,
    multimask_output=False,
)

后处理预测结果：

"""
Post-process the output of the model to get the masks, scores, and logits for visualization
"""
# convert the shape to (n, H, W)
if masks.ndim == 4:
    masks = masks.squeeze(1)


confidences = confidences.numpy().tolist()
class_names = labels

class_ids = np.array(list(range(len(class_names))))

labels = [
    f"{class_name} {confidence:.2f}"
    for class_name, confidence
    in zip(class_names, confidences)
]

输出结果可视化：

"""
Visualize image with supervision useful API
"""
img = cv2.imread(img_path)
detections = sv.Detections(
    xyxy=input_boxes,  # (n, 4)
    mask=masks.astype(bool),  # (n, h, w)
    class_id=class_ids
)

box_annotator = sv.BoxAnnotator()
annotated_frame = box_annotator.annotate(scene=img.copy(), detections=detections)

label_annotator = sv.LabelAnnotator()
annotated_frame = label_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels)
cv2.imwrite(os.path.join(OUTPUT_DIR, "groundingdino_annotated_image.jpg"), annotated_frame)
plt.figure(figsize=(9, 6))
plt.title(f"annotated_frame")
plt.imshow(annotated_frame[:,:,::-1])

mask_annotator = sv.MaskAnnotator()
annotated_frame = mask_annotator.annotate(scene=annotated_frame, detections=detections)
cv2.imwrite(os.path.join(OUTPUT_DIR, "grounded_sam2_annotated_image_with_mask.jpg"), annotated_frame)
plt.figure(figsize=(9, 6))
plt.title(f"annotated_frame")
plt.imshow(annotated_frame[:,:,::-1])

GroundingDINO 的 Box 效果，准确检测出 袜子 和 吉他，两类实体：

SAM2 的分割效果，如下：

转换成 COCO 数据格式：

def single_mask_to_rle(mask):
    rle = mask_util.encode(np.array(mask[:, :, None], order="F", dtype="uint8"))[0]
    rle["counts"] = rle["counts"].decode("utf-8")
    return rle

if DUMP_JSON_RESULTS:
    # convert mask into rle format
    mask_rles = [single_mask_to_rle(mask) for mask in masks]

    input_boxes = input_boxes.tolist()
    scores = scores.tolist()
    # save the results in standard format
    results = {
        "image_path": img_path,
        "annotations" : [
            {
                "class_name": class_name,
                "bbox": box,
                "segmentation": mask_rle,
                "score": score,
            }
            for class_name, box, mask_rle, score in zip(class_names, input_boxes, mask_rles, scores)
        ],
        "box_format": "xyxy",
        "img_width": w,
        "img_height": h,
    }
    
    with open(os.path.join(OUTPUT_DIR, "grounded_sam2_local_image_demo_results.json"), "w") as f:
        json.dump(results, f, indent=4)