Building upon our previous work, a Depth Object Detector (DOD) for object detection with integrated depth estimation, this study explores model size reduction through architecture optimization, while maintaining the focus on accuracy and efficiency for embedded systems. Drawing insights from an ablation study, we present a streamlined DOD architecture that achieves a significant reduction in size by approximately 61.46%, from the original 4.18 MB to 1.6 MB, while preserving a high degree of precision on the MinneApple dataset, demonstrating effectiveness in fruit detection tasks. This optimization makes the DOD even more suitable for deployment on resource-constrained devices with limited ROM size, such as microcontrollers commonly used in edge computing applications for the Internet of Things (IoT).