Spacecraft Object Detection: DrivenData

In this project, I used DrivenData's Pose Bowl: Spacecraft Detection and Pose Estimation Challenge as source data to practice using image detection models.

The object detection task involves drawing a bounding box around spacecrafts in a large dataset of images. See example below.

I built a tested a training loop locally and then ran it remotely on Kaggle's GPUs. In the end, I did not make a submission to the coding challenge. I found that full training cycles were taking too long, even on Kaggle's GPUs, and I wanted to move onto other projects. Still, this project was valuable practice using leading deep learning frameworks.

Skills and technologies used

• File handling (sys, pathlib for general file management and PIL for image files)
• Pre-processing and visualisation (numpy, pandas and matplotlib)
• Modelling (torch, torchvision, pytorch-lightning)
• Training (kaggle for training large datasets. This required managing CPU/ GPU usage using cuda in pytorch).
• Monitoring training process with tensorboard and testing submission files in a docker container.