Global-Wheat-Dection-1st-Place - GitLite

KO | EN

gitlite — search

#hacktoberfest #ai #golang #android #rust #deep-learning #linux #macos #nodejs #windows #kubernetes #pytorch #security #open-source

Global-Wheat-Dection-1st-Place

(★ 76)

1st place solution for Global Wheat Detection Challenge

파일 탐색기

최종 버전 다운로드 (.zip)

- trainset.csv
- - config.py
- - __init__.py
  - argmax_matcher.py
  - box_coder.py
  - box_list.py
  - faster_rcnn_box_coder.py
  - matcher.py
  - README.md
  - region_similarity_calculator.py
  - target_assigner.py
- __init__.py
- anchors.py
- bench.py
- efficientdet.py
- helpers.py
- loss.py
- download.sh
- gwd2020.png
- mixup.png
- mosaic.png
dataset.py
effdet_train.py
evaluate.py
evaluation.py
faster_rcnn_fpn_train.py
LICENSE
models.py
README.md
requirements.txt
utils.py
warmup_scheduler.py

# 설치 가이드

1. 코드 내려받기

git clone https://github.com/dungnb1333/Global-Wheat-Dection-1st-Place

깃허브에서 프로젝트 코드 전체를 내 컴퓨터로 내려받습니다.

cd Global-Wheat-Dection-1st-Place

방금 내려받은 프로젝트 폴더 안으로 이동합니다.

2. Python

쉬움 추천

사전 준비물

Git GitHub에서 프로젝트 코드를 내려받으려면 필요합니다.
Python 3 설치 시 'Add Python to PATH' 옵션을 꼭 체크하세요 (Windows).

$ pip install -r requirements.txt

requirements.txt 등에 명시된 파이썬 라이브러리를 설치합니다.

✅ 에러 메시지 없이 실행되고 터미널에 안내 문구가 출력되면 정상입니다.

이 레포의 README에 적힌 실제 명령어를 그대로 가져왔습니다.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

## Global Wheat Detection - 1st place solution

![Alt text](https://raw.githubusercontent.com/dungnb1333/Global-Wheat-Dection-1st-Place/main/images/gwd2020.png)

1st place solution of the [Global Wheat Detection Challenge](https://www.kaggle.com/c/global-wheat-detection)
The goal of this competition is to detect wheat heads in outdoor images collected from multiple regions around the world. This README provides a concise guide to reproducing the solution.

### Summary
Key techniques used in this solution:

* Custom Mosaic augmentation
* MixUp augmentation
* Heavy data augmentation
* EfficientDet
* Faster R-CNN with FPN
* Multi-scale ensemble using Weighted Boxes Fusion (WBF)
* Test-Time Augmentation (TTA)
* Pseudo-labeling

### Augmentations
* Custom mosaic augmentation
![Alt text](https://raw.githubusercontent.com/dungnb1333/Global-Wheat-Dection-1st-Place/main/images/mosaic.png)
* MixUp
* Heavy augmentation: RandomCrop, HorizontalFlip, VerticalFlip, ToGray, AdditiveGaussianNoise, GaussNoise, MotionBlur, MedianBlur, Blur, CLAHE, Sharpen, Emboss, RandomBrightnessContrast, HueSaturationValue

### Requirements
- Ubuntu 18.04 LTS
- CUDA 10.1
- cuDNN 7.5.1
- Python 3.7.6
- [nvidia-apex 0.1](https://github.com/NVIDIA/apex)

Setup environment
```
$ conda create -n wheat_env python=3.7.6
$ conda activate wheat_env
$ pip install -r requirements.txt
```

## DATASET
- Download the competition dataset from https://www.kaggle.com/c/global-wheat-detection/data
- Wheat 2017: https://plantimages.nottingham.ac.uk/
- Spike Wheat dataset: https://sourceforge.net/projects/spike-dataset/

Extract the datasets and prepare the labels.
```
$ cd dataset
$ unzip spike-wheat.zip
$ unzip wheat2017.zip
```
./dataset folder structure should be:
```
dataset
├── sample_submission.csv
├── test
│   ├── 2fd875eaa.jpg
│   ├── ...
├── train
│   ├── 00333207f.jpg
│   ├── ...
├── trainset.csv
├── wheat2017
│   ├── wheat2017_0001.jpg
│   ├── ...
├── wheat2017.csv
├── spike-wheat
│   ├── spike0000.jpg
│   ├── ...
├── spike-wheat.csv
```

### Model
* [EfficientDet-PyTorch](https://github.com/rwightman/efficientdet-pytorch) licensed under Apache 2.0, Copyright Ross Wightman
* [Faster RCNN FPN](https://github.com/pytorch/vision/tree/master/torchvision/models/detection) licensed under BSD 3-Clause
* 5 folds cross validation
* Optimizer: Adam with initial LR 5e-4 for EfficientDet and SGD with initial LR 5e-3 for Faster RCNN FPN
* LR scheduler: cosine-annealing
* Warm-up 20 epochs with trainset + wheat2017 dataset + spike wheat dataset -> train 80 epochs with trainset + wheat2017
* Pseudo labeling

### Train
- Train models
```
$ cd effdet-pretrained && bash download.sh && cd ..
$ python effdet_train.py --folds 0 1 2 3 4 --backbone ed7 --img-size 768 --batch-size 8 --workers 16 --use-amp True
$ python effdet_train.py --folds 1 3 --backbone ed7 --img-size 1024 --batch-size 4 --workers 16 --use-amp True
$ python effdet_train.py --folds 4 --backbone ed5 --img-size 512 --batch-size 20 --workers 16 --use-amp True
$ python effdet_train.py --folds 1 --backbone ed6 --img-size 640 --batch-size 12 --workers 16 --use-amp True
$ python faster_rcnn_fpn_train.py --folds 1 --backbone resnet152 --img-size 1024 --batch-size 20 --workers 16
```
- Evaluate models
```
python evaluate.py --folds 0 --network effdet --backbone ed7 --img-size 768 --batch-size 16 --workers 8
python evaluate.py --folds 1 --network fasterrcnn --backbone resnet152 --img-size 1024 --batch-size 16 --workers 8
```
### Results
| Network                  | image-size | Fold | Valid AP |
| :------------------------| :----------|:-----|:---------|
| EfficientDet-D7          | 768        | 0    | 0.710    |
| EfficientDet-D7          | 768        | 1    | 0.716    |
| EfficientDet-D7          | 768        | 2    | 0.707    |
| EfficientDet-D7          | 768        | 3    | 0.716    |
| EfficientDet-D7          | 768        | 4    | 0.713    |
| EfficientDet-D7          | 1024       | 1    | 0.718    | 
| EfficientDet-D7          | 1024       | 3    | 0.720    | 
| EfficientDet-D5          | 512        | 4    | 0.702    | 
| EfficientDet-D6          | 640        | 1    | 0.716    | 
| FasterRCNN-FPN-resnet152 | 1024       | 1    | 0.695    |

### Pseudo labeling
- **Base model**: EfficientDet-D6 (image size 640, Fold 1) — 0.716 validation AP
- **Round 1**:
  Train EfficientDet-D6 for 10 epochs using:
  - Training set + pseudo-labeled test set (generated from ensemble predictions)
  - Initialized from the base checkpoint  
  - Old test set: 0.7719 (Public LB) / 0.7175 (Private LB)
  - New test set: 0.7633 (Public LB) / 0.6787 (Private LB)
- **Round 2**:
  Continue training EfficientDet-D6 for 6 epochs using:
  - Training set + updated pseudo labels (from Round 1)
  - Initialized from Round 1 checkpoint  
  - Old test set: 0.7754 (Public LB) / 0.7205 (Private LB)
  - New test set: 0.7656 (Public LB) / 0.6897 (Private LB)

### Submission notebook

https://www.kaggle.com/code/nguyenbadung/gwd2020?scriptVersionId=39977641