Repository for Computer Vision project - 2024
Repository for Computer Vision project - 2024
– Objective: Developed a model for oriented object detection in aerial images to accurately predict object bounding boxes with arbitrary orientations.
– Process of Implementation: Implemented a single-stage anchor-free detection model using ResNet101, heatmaps, offset prediction, and box boundary-aware vectors for detecting and localizing objects.
– Outcome: Achieved 84.71% accuracy with 11 FPS detection speed, outperforming traditional methods on the DOTA dataset.
– Tools and Technology, Algorithms Used: Utilized Python, TensorFlow, ResNet101, and convolutional neural networks with heatmaps, smooth L1 loss, and binary cross-entropy loss algorithms.