模型描述 (Model Descriptio)

?Paper | ?️Video |

运行环境 (Operatig eviromet)

coda activate <your_ev>
git cloe https://github.com/IDEA-Research/GroudigDINO.git
cd GroudigDINO
pip istall -e .

代码范例 (Code example)

from modelscope.pipelies import pipelie
import torch
import os
from PIL import Image, ImageDraw, ImageFot
import umpy as p
import requests
from io import BytesIO

def plot_boxes_to_image(image_pil, tgt):
    H, W = tgt["size"]
    boxes = tgt["boxes"]
    labels = tgt["labels"]
    assert le(boxes) == le(labels), "boxes ad labels must have same legth"

    draw = ImageDraw.Draw(image_pil)
    mask = Image.ew("L", image_pil.size, 0)
    mask_draw = ImageDraw.Draw(mask)

    # draw boxes ad masks
    for box, label i zip(boxes, labels):
        # from 0..1 to 0..W, 0..H
        box = box * torch.Tesor([W, H, W, H])
        # from xywh to xyxy
        box[:2] -= box[2:] / 2
        box[2:] += box[:2]
        # radom color
        color = tuple(p.radom.radit(0, 255, size=3).tolist())
        # draw
        x0, y0, x1, y1 = box
        x0, y0, x1, y1 = it(x0), it(y0), it(x1), it(y1)

        draw.rectagle([x0, y0, x1, y1], outlie=color, width=6)
        # draw.text((x0, y0), str(label), fill=color)

        fot = ImageFot.load_default()
        if hasattr(fot, "getbbox"):
            bbox = draw.textbbox((x0, y0), str(label), fot)
        else:
            w, h = draw.textsize(str(label), fot)
            bbox = (x0, y0, w + x0, y0 + h)
        # bbox = draw.textbbox((x0, y0), str(label))
        draw.rectagle(bbox, fill=color)
        draw.text((x0, y0), str(label), fill="white")

        mask_draw.rectagle([x0, y0, x1, y1], fill=255, width=6)

    retur image_pil, mask

pipe = pipelie('tiefag-groudig-dio-task',
                model='goldsj/GroudigDINO', model_revisio='v1.0.0')


iput_image_url = "https://ofasys.oss-c-zhagjiakou.aliyucs.com/data/coco/2014/val2014/COCO_val2014_000000222628.jpg"
respose = requests.get(iput_image_url)
iput_image_file =BytesIO(respose.cotet)

data = {
    'img': iput_image_file,
    'captio': "perso"
}
boxes_filt, pred_phrases = pipe(data)

# visualize pred

image_pil = Image.ope(iput_image_file).covert("RGB")

size = image_pil.size
pred_dict = {
    "boxes": boxes_filt,
    "size": [size[1], size[0]],  # H,W
    "labels": pred_phrases,
}
image_with_box = plot_boxes_to_image(image_pil, pred_dict)[0]
image_with_box.save("./pred.jpg")

Citatio

If you fid our work helpful for your research, please cosider citig the followig BibTeX etry.

@iproceedigs{ShilogLiu2023GroudigDM,
  title={Groudig DINO: Marryig DINO with Grouded Pre-Traiig for Ope-Set Object Detectio},
  author={Shilog Liu ad Zhaoyag Zeg ad Tiahe Re ad Feg Li ad Hao Zhag ad Jie Yag ad Chuyua Li ad Jiawei Yag ad Hag Su ad Ju Zhu ad Lei Zhag},
  year={2023}
}

模型描述 (Model Description) ?Paper | ?️Video | 运行环境 (Operating environment) conda activate