Pose-Bowl-Spacecraft-Detection

2nd place solution for Pose Bowl: Spacecraft Detection Challenge

General

• Competition Purpose: identify the boundaries of generic spacecraft in photos.
• Type: Object detection
• Host: NASA
• Platform: Drivendata
• Competition link: https://www.drivendata.org/competitions/260/spacecraft-detection/
• Placement: 2nd (2/651) + speed bonus prize
• User Name: dungnb
• Solution: SolutionDocumentation_SpacecraftDetection_DungNB.pdf

The key of solution is based on synthetic data generation and yolov8 model.

System

• Operating System: Ubuntu 22.04
• Nvidia Driver Version: 545.23.08
• Cuda: Version 11.3
• GPU: 1xNvidia A100 40GB
• RAM: 64GB
• Train duration: yolov8s 1280 (~60 hours), yolov8n 1280 (~24hours)
• Inference duration: yolov8s 1280(1 hour 40 minutes), yolov8n 1280(46 minutes)

Environment

conda create -n venv python=3.9.6
conda activate venv
conda install pytorch=1.12.1 torchvision=0.13.1 torchaudio=0.12.1 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Data preparation

• Download competition dataset and extract to folder ./dataset/drivendata

• Download public dataset in paper at repo or google drive to folder ./dataset/satellite_external
  Note: This data has been confirmed at discussion

• dataset structure should be ./dataset/dataset_structure.txt

And run following scripts

cd src
python prepare_drivendata_spacecraft.py
python prepare_external_data.py
python generate_synthetic_data.py

Train model

Change line 1 in file ./src/spacecraft_data.yaml to absolute path of folder ./dataset/ in your system

And run following scripts

cd src
rm -rf runs
python train.py

Export model to openvino

Change path of trained model in file ./src/export_to_openvino.py

cd src
python export_to_openvino.py

Inference

Step by step to generate submission file

Result

	Public LB	Private LB	Runtime
yolov8s 1280	0.9285	0.9226	1 hour 40 minutes
yolov8n 1280	0.9173	0.9098	46 minutes