01. The Challenge
High-resolution aerial and drone imagery presents significant computer vision challenges due to extreme object scale variance, dense clustering, occlusions, and diverse lighting conditions across geographic regions.
The objective was to evaluate, benchmark, and optimize modern deep learning object detection models (YOLOv8 variants) to perform real-time inference on geospatial datasets with minimal false positives.
02. Vision Pipeline Flow
│
[ Slicing & Data Augmentation (SAHI) ]
│
[ YOLOv8 Backbone & Feature Pyramid ]
│
[ Bounding Box Regression & Classification ]
03. Implementation
Data Preprocessing & SAHI
Employed Sliced Aided Hyper Inference (SAHI) to divide ultra-high-resolution aerial images into overlapping patches, ensuring small targets maintain high feature resolution during model evaluation.
Model Training & Hyperparameter Tuning
Trained YOLOv8 Nano and Medium models using custom learning rate scheduling, Mosaic augmentation, and automated anchor tuning over 100 epochs on GPU clusters.
04. Key Results
- 01. Achieved 91.4% mAP@0.5 on challenging small-object aerial benchmarks.
- 02. Real-time inference speed of 45 FPS on standard edge GPU hardware.
- 03. Formed the foundation of the Software Engineering Bachelor's thesis with distinction.