VRDL-Final Project

NCTU Selected Topics in Visual Recognition using Deep Learning Final Project.

Global Wheat Detection

The topic is from Kaggle. It is about detecting wheat from images.

Hardware

The following specs were used to create the original solution.

• Ubuntu 18.04 LTS
• Intel(R) Core(TM) i7-6700 CPU @ 3.40 GHz
• NVIDIA GeForce GTX TITAN X

Installation

1. Using Anaconda is strongly recommended. {envs_name} is the new environment name which you should assign.

   conda create -n {envs_name} python=3.7
   conda activate {envs_name}

2. Install PyTorch and torchvision following the official instructions, e.g.,

   conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch

   Note: Make sure that your compilation CUDA version and runtime CUDA version match. You can check the supported CUDA version for precompiled packages on the PyTorch website.
3. Install requirements.

   pip install -r requirements.txt

Dataset Preparation

You can download the dataset here.

You can also download the pretrained models from EfficientDet Pytorch.

After downloading the data and pretrained models, the data directory is structured as:

+- input
    +- efficientdet
        +- efficientdet_d0-d92fd44f.pth
        +- efficientdet_d1-4c7ebaf2.pth
        ...
    +- global-wheat-detection
        +- train
            +- 0a3cb453f.jpg
            +- 0a3ff84a7.jpg
            ...
        +- test
            +- 2fd875eaa.jpg
            +- 51b3e36ab.jpg
            ...
        train.csv
        sample_submission.csv
+- EfficientDetTool
+- omegaconf
+- weightedboxesfusion
train.py
make_submission.py

Training

You can train the model by following:

python3 train.py

Testing

You can test the model and make a csv submission file by following:

python3 make_submission.py

Thanks

Big thank for Alex Shonenkov

[Training] EfficientDet

[Inference] EfficientDet

WBF approach for ensemble

[WBF over TTA][Single Model] EfficientDet