xView3_3rd_place_solution

3rd place solution for xView3 challenge https://iuu.xview.us/

For a detailed description of the method, please refer to "./doc/xView3_competition_third_place_solution.pdf".

The submission docker image com be download from docker hub.

docker pull sanxia04/heatmap:v4.1

Getting Started

Clone the repository:

git clone https://github.com/Tumenn/xView3_3rd_place_solution.git

We use Python 3.6 and PyTorch 1.9.0 in our implementation, please install dependencies:

conda create -n xview python=3.6
conda activate xview
conda install pytorch=1.9.0 cudatoolkit=11.1 -c pytorch
conda install GDAL==3.0.2
pip install -r requirements.txt

Data Preparation

First, please download the xview3 competition dataset and unzip it in the folder ./data/raw_data/ .

The structure of the data folder should be:

├── data
    ├── raw_data                   # xview3 competition dataset
        ├── train                  # train dataset
        |   ├── 00a035722196ee86t
        |   |   ├── VH_dB.tif
        |   |   ├── VV_dB.tif
        |   |── 0aba50e606eddffet
        |   |── ...
        ├── valid                  # valid dataset
        |   ├── 0d8ed29b0760dc59v
        |   |── ...
        ├── test                   # test dataset
        |
        ├── train.csv              # train label
        ├── valid.csv              # valid label

Run the following command to process the xView3 competition dataset :

cd data_preprocess
python convert_8bit,py
python detection.py
python classfication.py

After process, the structure of the data folder should be:

├── data
    ├── raw_data                   # xview3 competition dataset
    ├── split_txt                  # store k fold image path
    ├── train_8bit                 # train images in 8 bit
    ├── train_class_crop           # train crops for classification model
    ├── valid_8bit                 # valid images in 8 bit
    ├── valid_class_crop           # valid crops for classification model
    ├── valid_detection_crop       # valid crops for detection model

Training

We use the RTX 3090 GPU to train the network.

Download the pretrained model of HRNet and place it under the './data/ckpt/' folder.

Run the following command to train the classification model on one GPU:

cd src
sh train_class.sh

Run the following command to train the detection model on one GPU:

sh train_detect.sh

If you have multiple GPUs, you can change the parameters in .sh file to train multiple models parallelly.

Inference

Run the following command to inference one test scene:

cd inference
sh run_inference.sh ../../data/raw_data/test 0157baf3866b2cf9v ../../data/raw_data/predict.csv

The shell script takes 3 positional arguments The three positional arguments are:

1. --image_folder which is the path to the data directory.
2. --scene_ids which is the list of the xView3 scene identifier for which you wish to run inference.
3. --output which is the path to the prediction output filename in CSV format.