Autonomous vehicles and robots require a three-dimensional understanding of their surroundings to navigate effectively. One critical aspect that contributes to this 3D comprehension is object detection. Existing methods for 3D object detection generally rely on three different types of input data: LiDAR, stereo vision, and monocular (single-camera) vision. Monocular methods have consistently garnered attention due to their simpler implementation and lower costs. However, the absence of depth information in monocular images presents unique challenges, particularly in depth estimation. Given the crucial role of depth estimation in monocular 3D object detection, this research proposes a novel approach to enhance depth estimation based on input image style transfer. This proposed method resulted in a relative improvement of 16% in the a1 metric and a relative improvement of 44% in the AbsRel metric.Further in this research, two additional methods are proposed to improve monocular 3D object detection results. The first proposed method involves using unlabeled images from the dataset with weak labels for pre-training and initial network weighting. This approach led to a relative improvement of 8.25% in the mean average precision (mAP) eva‎luation metric. The second proposed method aims to improve monocular 3D object detection by enforcing the network to consider the depth-to-height ratio of objects according to geometric relationships within the image. This method, by adding an auxiliary regression head to the base network, resulted in a relative improvement of 9.7% in the mean average precision for car 3D detection.

نادر كريمي , شادرخ سماوي

محمدرضا احمدزاده , جلال ذهبي