🦖 DINO Practice: Running Object Detection with Pretrained Models - DINO 실습: 모델을 받아 직접 객체 탐지 해보기!

(English) 🦖 DINO Practice: Running Object Detection with Pretrained Models!

In the previous post, I studied the core ideas behind DINO!!!
Now, let’s dive into actually installing the DINO model from GitHub and running object detection on our own images~!

• Starting with the result, as always!!
• Just like DETR, it detects multiple objects in the image and visualizes them!!
• Let’s walk through each step together in Python code!
• This time, there’s no simple pip install – we need to download and build from the GitHub repo. Let’s get started!! 😄

---

🔗 1. Clone the DINO repository from GitHub

🔗 Official GitHub Repo

A very well-organized GitHub repository.
Huge thanks to the author!! 🙏

---

📦 2. Create a virtual environment and install dependencies

conda create --name DINO python=3.9
conda activate DINO

sudo apt update && sudo apt install -y build-essential python3-dev
pip install cython

conda install -c conda-forge libstdcxx-ng
conda install -c pytorch pytorch torchvision

pip install -r requirements.txt

---

⚙️ 3. Build and test the model code

cd models/dino/ops
python setup.py build install

# Unit test - should print all True if successful
python test.py

cd ../../..

---

🗂️ 4. Prepare the COCO2017 dataset

Organize the COCO dataset as follows:

COCODIR/
├── train2017/
├── val2017/
└── annotations/
    ├── instances_train2017.json
    └── instances_val2017.json

📥 5. Download Pretrained Model

You can download the pretrained DINO-4scale, Swin-L (36 epochs) model from the following Google Drive link:

🔗 https://drive.google.com/drive/folders/1qD5m1NmK0kjE5hh-G17XUX751WsEG-h_?usp=sharing

---

🧠 6. Load the Model and Run Inference

Import Packages

import os, sys
import torch, json
import numpy as np

from main import build_model_main
from util.slconfig import SLConfig
from datasets import build_dataset
from util.visualizer import COCOVisualizer
from util import box_ops

from PIL import Image
import datasets.transforms as T

Load COCO Class ID-to-Name Mapping

with open('{your_path1}/DINO/util/coco_id2name.json') as f:
    id2name = json.load(f)
    id2name = {int(k):v for k,v in id2name.items()}

Load Model Configuration and Checkpoint

model_config_path = "{your_path1}/DINO/config/DINO/DINO_4scale.py"
model_checkpoint_path = "{your_path2}/checkpoint0033_4scale.pth"

args = SLConfig.fromfile(model_config_path) 
args.device = 'cuda' 
model, criterion, postprocessors = build_model_main(args)

checkpoint = torch.load(model_checkpoint_path, map_location='cpu')
model.load_state_dict(checkpoint['model'])
_ = model.eval()

Load and Preprocess the Image

image = Image.open("{image_path}/catch_rugby_ball_001480.jpg").convert("RGB")

transform = T.Compose([
    T.RandomResize([800], max_size=1333),
    T.ToTensor(),
    T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
image, _ = transform(image, None)

Run Inference

output = model.cuda()(image[None].cuda())
output = postprocessors['bbox'](output, torch.Tensor([[1.0, 1.0]]).cuda())[0]

---

🖼️ 7. Visualize the Results

thershold = 0.3  # set confidence threshold

vslzr = COCOVisualizer()

scores = output['scores']
labels = output['labels']
boxes = box_ops.box_xyxy_to_cxcywh(output['boxes'])
select_mask = scores > thershold

box_label = [id2name[int(item)] for item in labels[select_mask]]
pred_dict = {
    'boxes': boxes[select_mask],
    'size': torch.Tensor([image.shape[1], image.shape[2]]),
    'box_label': box_label
}

vslzr.visualize(image, pred_dict, savedir=None, dpi=100)

And just like that – boom! The detection works beautifully! 🙌

---

✅ Wrap-Up

At first, I was a bit nervous about using models not yet uploaded to timm or HuggingFace,
but once I got into it, everything was quite manageable!
Especially in this case, the DINO repo was incredibly well organized,
which made the process even smoother.

I’ll continue experimenting with more models and sharing my findings. Stay tuned! 🚀

---

(한국어) 🦖 DINO 실습: 모델을 받아 직접 객체 탐지 해보기!

지난 포스팅 에서 DINO의 원리에 대하 공부해보았습니다!!!
이젠, DINO 모델을 git repo로부터 직접 설치하여 객체 탐지(Object Detection)을 진행해 보겠습니다~!

• 오늘도도 결론부터!!!
• DETR과 동일하게 이미지에서 탐지된 여러 객체들을 찾아서 보여줍니다!!
• 함께, 파이썬 코드로 그 과정을 알아보아요!!
• 이번에는 간단한 모델이 없어 git 에서 모델을 받고 설치해야합니다~ 잘 따라해보세요!!^^

🔗 1. GIT 저장소에서 DINO 모델 받기!!

🔗 공식 GitHub 저장소

아주 잘 정리된 git repo입니다~!
작성자분께 감사드립니다!!^^

---

📦 2. 가상환경 및 의존성 설치

conda create --name DINO python=3.9
conda activate DINO

sudo apt update && sudo apt install -y build-essential python3-dev
pip install cython

conda install -c conda-forge libstdcxx-ng
conda install -c pytorch pytorch torchvision

pip install -r requirements.txt

---

⚙️ 3. 모델 코드 컴파일 및 테스트

cd models/dino/ops
python setup.py build install

# 유닛 테스트 - 모두 True가 출력되면 성공
python test.py

cd ../../..

---

🗂️ 4. COCO2017 데이터셋 준비

COCO2017 데이터를 아래 구조로 정리합니다:

COCODIR/
├── train2017/
├── val2017/
└── annotations/
├── instances_train2017.json
└── instances_val2017.json

---

📥 5. 사전 학습된 모델 다운로드

DINO-4scale, Swin-L, 36 epoch 기준 모델은 아래 Google Drive 링크에서 다운로드합니다:

🔗 https://drive.google.com/drive/folders/1qD5m1NmK0kjE5hh-G17XUX751WsEG-h_?usp=sharing

---

🧠 6. 모델 로드 및 이미지 예측

패키지 임포트

import os, sys
import torch, json
import numpy as np

from main import build_model_main
from util.slconfig import SLConfig
from datasets import build_dataset
from util.visualizer import COCOVisualizer
from util import box_ops

from PIL import Image
import datasets.transforms as T

COCO 클래스 ID 매핑 파일 로드

with open('{알맞은위치1}/DINO/util/coco_id2name.json') as f:
    id2name = json.load(f)
    id2name = {int(k):v for k,v in id2name.items()}

모델 구성 및 체크포인트 로드

model_config_path = "{알맞은위치1}/DINO/config/DINO/DINO_4scale.py"
model_checkpoint_path = "{알맞은위치2}/checkpoint0033_4scale.pth"

args = SLConfig.fromfile(model_config_path) 
args.device = 'cuda' 
model, criterion, postprocessors = build_model_main(args)

checkpoint = torch.load(model_checkpoint_path, map_location='cpu')
model.load_state_dict(checkpoint['model'])
_ = model.eval()

이미지 불러오기 및 전처리

image = Image.open("{이미지위치}/catch_rugby_ball_001480.jpg").convert("RGB")

transform = T.Compose([
    T.RandomResize([800], max_size=1333),
    T.ToTensor(),
    T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
image, _ = transform(image, None)

이미지 예측

output = model.cuda()(image[None].cuda())
output = postprocessors['bbox'](output, torch.Tensor([[1.0, 1.0]]).cuda())[0]

---

🖼️ 7. 결과 시각화

thershold = 0.3  # 임계값 설정

vslzr = COCOVisualizer()

scores = output['scores']
labels = output['labels']
boxes = box_ops.box_xyxy_to_cxcywh(output['boxes'])
select_mask = scores > thershold

box_label = [id2name[int(item)] for item in labels[select_mask]]
pred_dict = {
    'boxes': boxes[select_mask],
    'size': torch.Tensor([image.shape[1], image.shape[2]]),
    'box_label': box_label
}

vslzr.visualize(image, pred_dict, savedir=None, dpi=100)

하면~~ 짠 아래와 같이!! 잘 detection 하네요~~

---

✅ 마무리

timm, huggingface 등에 모델이 업데이트 되지 않으면 만만치 않겠다라는 두려움이생기지만~,
막상 해보면 다 할수 있습니다!
특히 이번 DINO는 모델 업로드한 작성자가 깔끔하게 정리되어있어 더욱 쉽게할수 있었습니다!!

앞으로도 여러 모델을 테스트해보겠습니다!!