Shiina-Mahiru/Pet-Mischief-Detection

COMP4423 – Computer Vision

Project - The Pet Mischief Detector

LI Jianxi 24109859d

Zhang Xulu 24113522d

CHENG Keyan 24110027d

Task 1: Problem Definition & Data Curation

1.1 Define Classes and Rules

Pets (2 classes):

• cat
• dog

Objects (5 classes):

• cup: Easy target for cats to push
• laptop: Dogs might chew, cats might sit on keyboard
• potted_plant: Cats love digging in soil
• vase: Fragile, cats love knocking over
• couch: Added this for "safe scenario" detection

Total: 7 classes

1.2 Curate the Dataset

Curate dataset from COCO 2017. We mix the original train and val together.

Task 2: Dataset Splitting & Formatting

• Split the Data: Programmatically split the curated dataset into training, validation, and test sets (e.g., 80%/10%/10%). Ensure this split is reproducible by using a fixed random seed RANDOM_SEED = 42.

• Format for Training: Convert the dataset into the format required by the chosen detection framework (Here is YOLOv8n). Document this process and provide any conversion scripts.

Task 3: Training/Fine-tuning a Vision Model

• Select and Train a Model: Choose a vision model and train it on your custom dataset. Here we choose YOLOv8n.

Task 4: Mischief System Implementation & Evaluation

Part A: Quantitative Detector Evaluation

Part B: Qualitative System Evaluation