Abstract
Deep learning models for object detection and segmentation need large labeled datasets. Manual labeling is slow, costly, and often has errors. This paper introduces an automated pipeline to generate synthetic segmentation datasets using 3D rendering and contour-based mask extraction. It removes the need for manual annotation. The method uses Blender with the Cycles engine to render 3D object models from multiple viewpoints on plain backgrounds. And, OpenCV is then used to remove the background and extract contours to create binary segmentation masks compatible with YOLO. The pipeline of this study was run on Google Colab with a T4 GPU. It generated 360 images per object with automated mask creation. A YOLOv5s-seg model trained on a dataset of two objects (720 images) for 50 epochs achieved 0.942 mAP@0.5 for detection and$0.887 \text{mAP} {@} 0.5$for segmentation. Precision was 0.94, and recall was 0.91. White backgrounds gave about 98% mask accuracy (IoU), while black backgrounds gave 89%. The system kept a real-time inference speed of 20ms per frame and generalized well to unseen viewpoints. This approach shows that synthetic data generation can create high-quality datasets without manual labeling. It reduces preparation time to zero manual steps and provides a scalable solution for object-centered computer vision tasks where collecting real data is hard or expensive.