I built AERIS as an end-to-end computer-vision pipeline for aerial Intelligence, Surveillance, and Reconnaissance (ISR), using real satellite imagery to detect military-relevant assets. This project combines detection, scene handling, target prioritization, and a multimodal architecture.
I fine-tuned YOLOv8m-OBB on DOTA v1 and achieved mAP50 = 0.909 across 2,000 validation scenes.
My YOLOv8m-OBB result on DOTA v1 was mAP50-95 = 0.763.
I built AERIS, an end-to-end computer-vision pipeline for aerial ISR, trained on 54,087 real satellite images across 16 classes of military-relevant assets.
I use overlap tiling for giga-pixel satellite scenes because a T-72 tank at 30cm GSD collapses to <1 pixel if a scene is naively resized to 640².
I use data-driven oriented bounding boxes because 17.6% of DOTA objects have aspect ratio >3:1.
For elongated DOTA objects, axis-aligned boxes waste 50%+ of their area on background.
On a free Kaggle T4, YOLOv8n-OBB ran at 63 FPS at 640².
On a free Kaggle T4, YOLOv8m ran at ~10 FPS at 1024².
I designed operator-tunable target prioritization so operators can re-rank targets for air / maritime / ground mission modes without retraining, with per-signal audit trails.
My multi-modal EO / SAR / IR fusion backbone has 3.76M params and uses gated per-pixel attention; it is architected and ready for paired-data training.
My per-class detection results included plane AP50 = 0.992.
My per-class detection results included ship AP50 = 0.976.
My per-class detection results included helicopter AP50 = 0.972.
I completed this work on Kaggle’s free tier, with no proprietary data and no expensive hardware.
I am open to research / engineering roles in defense computer vision.