Self-Supervised Monocular Depth Estimation Using HOG Feature Prediction

Accurate depth estimation from monocular images continues to pose a formidable challenge owing to boundary blurring, illumination variations, and occlusions encountered in traditional monocular depth estimation methods. In response, this paper introduces a fresh approach to self-supervised monocular depth estimation leveraging HOG (Histogram of Oriented Gradients) feature prediction. The method aims to mitigate the aforementioned challenges by preserving fine details at object boundaries and improving prediction accuracy. Central to our methodology is the introduction of a HOG feature prediction module. This module meticulously extracts HOG feature vectors from input image characteristics while emphasizing the retention of crucial boundary information. Furthermore, it strategically enhances the encoder's downward output features, thereby refining the depth estimation process. Our proposed approach has been thoroughly assessed through comprehensive experimental evaluations on the extensively used KITTI datasets, affirming its efficacy. The results underscore its superiority when compared to prevailing mainstream methodologies, particularly in accurately predicting fine-grained depth details along object edges. The proposed framework exhibits robustness against boundary blurring, illumination variations, and occlusions, offering promising advancements in monocular depth estimation techniques.