합성데이터 기반 객체 탐지 AI 경진대회

합성 데이터를 활용한 자동차 탐지 AI 모델 개발

• 합성데이터란 실제 환경에서 수집되거나 측정되는 것이 아니라 디지털 환경에서 생성되는 데이터셋으로, 최근 방대한 양질의 데이터셋이 필요해짐에 따라 그 중요성이 대두되고 있습니다.

• 합성 데이터는 데이터 라벨링 작업을 위한 2배 이상의 시간 절약과 10배 가까운 비용을 절감하게 하고, 자동화를 바탕으로 정확한 라벨링의 데이터 그리고 정확한 AI 모델 개발을 위한 데이터의 다양화를 가능하게 합니다.

• 학습용 합성데이터를 활용하여 자동차 탐지를 수행하는 AI 모델을 개발해야 합니다.

Project structure

Synthetic-Data-Object-Detection
├─ .gitignore
├─ archive  # implementation pytorch 
├─ data  
│  ├─ raw
│  │  └─ raw data
│  ├─ ensemble
│  │  └─ ensemble data
│  └─ submission
├─ models
│  └─ model file
├─ mmdetection
│  ├─ configs  
│  │  ├─ _base_  
│  │  └─ visol  # model config
│  │      └─ model config file.py
│  └─ mmdet
├─ mmdetection3.x/
├─ inference.py # Test Inference
├─ grad_cam.py  # GradCam
├─ ensemble.py  # weighted boxes fusion
└─ README.md

Getting Started

Python 3.8.10 mmdetection 2.x

git clone https://github.com/Myungbin/Synthetic-Data-Object-Detection.git

python -m pip install --upgrade pip
python -m pip install -r requirements.txt

cd mmdetection
python .\tools\train.py .\configs\visol\{config_file.py}
python .\tools\test.py '{config_file_path}' '{model_result_path}' --format-only --out '{result_path}'

Data

Dataset Link
The project utilizes a custom dataset for training and evaluation. The dataset consists of labeled images with bounding box annotations for each object of interest. The dataset is not included in this repository and needs to be prepared separately.

Experiment

The default augmentations used were Resize, Flip, and Normalize.
Faster R-CNN, EfficientDet, Swin Transformer, Libra R-CNN, and YOLO were used in the experiment. You can check the other experiments in the "finished_experiments" folder. In the end, Cascade R-CNN models was used.

Model	Backbone	Depth	Augmentation	mAp
Cascade R-CNN	SwinT	-	Mixup, Cutout	0.89
Libra R-CNN	Resnest	200	Mixup, Cutout, AutoAugment, PhotoMetricDistortion	0.89
Faster R-CNN	ResNeXt	101	Mixup	0.91
Faster R-CNN	ResNeSt	200	Mixup	0.93
Faster R-CNN	ResNeSt	101	Mixup, Cutout	0.95
Cascade R-CNN	ResNeSt	200	Mixup, Cutout, AutoAugment, PhotoMetricDistortion	0.98
...	...	...	...	...

Ensemble

Weighted-Boxes-Fusion on the results of 8 models.

pretrain model

You can check the pretrained model weights google drive

Result

Public mAp 0.9964 Private mAP 0.99403

To reproduce the results of a public mAP of 0.9964 and a private mAP of 0.99403, follow these steps

1. v7.py must be trained and inferred using a seed of 2023 and the NVIDIA RTX 4090 GPU. Additionally, this file utilizes mmdetection 3.x, it should be executed based on the following document.
2. In the final folder, v3.py must be trained and inferred using a seed of 1927851590 and the NVIDIA RTX 3090 GPU.
3. For all other cases must be trained and inferred using a seed of 378452678 and the NVIDIA A100 GPU.
4. Finally, You must ensemble the inference results of each model using Weighted-Boxes-Fusion. The weights for models v1 to v8 are [0.05, 0.05, 0.05, 0.05, 0.2, 0.2, 0.2, 0.2].

You can find detailed information about the experimental results through the PowerPoint presentation.

Development Environment

OS: Window11
CPU: Intel i9-11900K
RAM: 128GB
GPU: NVIDIA GeFocrce RTX3090 & RTX4090 & A100 & V100

Host

• 주최 : 비솔(VISOL)
• 주관 : 데이콘(DACON)
• 기간 : 2023.05.08 ~ 2023.06.19 09:59
  Competition Link