Dockerized Dataset Prepocessing -> Training Pipeline

[sah4jpatel]

input:

• a path directory of unlabelled images OR a labelled data in COCO format
• a flag to specify whether labelled or unlabeled input
• path to auxiliary input files (model files most likely)

intermediate steps:

• set up a python3.10 environment with the "ultralytics" package installed
• also install the python3.10 rknn-toolkit2 wheel

run process:

• if unlabeled: run the dataset autolabeller.py file that I will provide to you
• else: run the dataset augmentation.py script that I will provide to you
• then, pass the resulting datasets (you will get 2) into a ultralytics yolov11n training script (once per dataset)
• finally, convert the final "best.pt" trained models for each run in ".rknn" format using ultralytics' export function

output:

• the trained "balloons.rknn" model and metadata file
• the trained "goals.rknn" model and metadata file