Global Wheat Detection - 1st place solution

This repo contains the source code of the 1st place solution for Global Wheat Detection Challenge. In this competition, you’ll detect wheat heads from outdoor images of wheat plants, including wheat datasets from around the globe. Below you can find a outline of how to reproduce my solution.

Summary

• Custom mosaic data augmentation
• MixUp
• Heavy augmentation
• EfficientDet
• Faster RCNN FPN
• Ensemble multi-scale model: Weighted-Boxes-Fusion
• Test time augmentation(HorizontalFlip, VerticalFlip, Rotate90)
• Pseudo labeling

Augmentations

• Custom mosaic augmentation
• MixUp
• Heavy augmentation: RandomCrop, HorizontalFlip, VerticalFlip, ToGray, AdditiveGaussianNoise, GaussNoise, MotionBlur, MedianBlur, Blur, CLAHE, Sharpen, Emboss, RandomBrightnessContrast, HueSaturationValue 2 examples mixup+mosaic+augmentation:

Requirements

• Ubuntu 18.04 LTS
• CUDA 10.1
• CuDNN 7.5.1
• Python 3.7.6
• python packages

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

• nvidia-apex 0.1

DATASET

• download competition dataset at link then extract to ./dataset folder
• Wheat 2017
• Spike-wheat dataset
• extract external dataset and label

$ 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 licensed under Apache 2.0, Copyright Ross Wightman
• Faster RCNN FPN 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 all models from scratch

• 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

Performance

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: EfficientDet-d6 image-size 640 Fold1 0.716 Valid AP
• Round1: Train EfficientDet-d6 10 epochs with trainset + hidden testset (output of ensembling), load weight from base checkpoint
  Result: [old testset] 0.7719 Public LB/0.7175 Private LB and [new testset] 0.7633 Public LB/0.6787 Private LB
• Round2: Continue train EfficientDet-d6 6 epochs with trainset + hidden testset (output of pseudo labeling round1), load weight from pseudo labeling round1 checkpoint
  Result: [old testset]0.7754 Public LB/0.7205 Private LB and [new testset]0.7656 Public LB/0.6897 Private LB

Kaggle kernel

Final submission