A Unified Roadmap of Deep Convolutional Neural Networks for Object Detection

Hrishi Rakshit, Pooneh Bagheri Zadeh, Akbar Sheikh Akbari

Although deep convolutional neural networks (DCNNs) have transformed object detection by automating the extraction of reliable feature representations, researchers find it challenging to monitor small improvements and pinpoint unresolved issues due to the quick spread of various architectures. This research traces the development of DCNN designs from early sliding-window constraints to complex modern frameworks, offering a thorough, comparative synthesis of the structural evolution of the discipline. This work’s main contribution is the provision of a unified roadmap that methodically assesses the core ideas, advantages, and intrinsic trade-offs – like the trade-off between speed and accuracy – found in two-stage and single-stage detectors. This synthesis enables researchers to effectively discover gaps in current literature by combining a comprehensive evaluation of cutting-edge backbone networks and benchmark datasets with an in-depth investigation of fundamental CNN mechanics. In the end, this study eliminates the need for thorough assessments of individual studies by acting as an essential resource for choosing the best designs for certain computer vision applications.