Code for object detection adapting EfficientDet.

• Team: The Fastest Kokushimuso in Hsintsu
• Competetion: Kaggle Global Wheat Detection ECCV 2020
• Method: EfficientDet

• Ubuntu 16.04 LTS

• Most of the code based on shonenkov's notebook, we appreciate his awesome work.

1. Installation
2. Dataset Preparation
3. Training

First, clone our wheat-detection repo

$ https://github.com/osinoyan/wheat-detection
$ cd wheat-detection

All requirements should be detailed in requirements.txt. Using Anaconda is strongly recommended.

conda create -n wheat python=3.6
source activate wheat
pip install -r requirements.txt

Download the dataset from kaggle and unzip global-wheat-detection.zip under the directory data/. Make sure to place the data like below:

    wheat-detection/
        +- data/
        |   +- sample_submission.csv
        |   +- train.csv
        |   +- test/
        |   |   +- 2fd875eaa.jpg
        |   |   +- ...
        |   +- train/
        |   |   +- 0a3cb453f.jpg
        |   |   +- ...
        ... ...

We briefly provide the instructions to train the model

Unfortunately, you have to download the weights manually. Download efficientdet_d5-ef44aea8.pth from This link and make sure to place it under the directory efficientdet_model/ like below:

    wheat-detection/
        +- efficientdet_model/
        |   +- efficientdet_d5-ef44aea8.pth
        +- data/
        +- timm-efficientdet-pytorch-revised/
        +- weighted-boxes-fusion/
        |
        ...

Just run this:

python train.py