Undergraduate Thesis — Flagship Research
Deformable convolutions, tested where it matters: real traffic.
Standard convolutions assume rigid geometry — but vehicles occlude, pedestrians lean, and CCTV angles distort. This thesis ran a controlled ablation study of Deformable Convolutional Network variants across YOLOv8 model sizes, using real-world CCTV footage for vehicle and pedestrian detection and tracking.
Nine configurations were trained and evaluated end to end: dataset preparation, training harnesses, metric evaluation, and a critical comparison of how model capacity and DCN placement affect performance.
Model capacity and DCN placement interact — gains depend on where deformable modules sit in the network.
Full-neck DCNv2 on YOLOv8m improved both precision and recall over the baseline configuration.
Real-world CCTV footage exposes deployment gaps that curated benchmarks hide.
- Python
- PyTorch
- YOLOv8
- OpenCV
- DCNv2
Ablation matrix
6 configs · simplified
| Model | Baseline | DCN · partial | DCNv2 · full neck |
|---|---|---|---|
| YOLOv8n | reference | evaluated | evaluated |
| YOLOv8m | reference | evaluated | Strongest |
Directional findings only — precision and recall improved over the baseline for the strongest configuration. Full metrics live in the thesis document.
Winning configuration
YOLOv8m
+ DCNv2 · full neck
The strongest performer across the study — lifting both precision and recall over the baseline.