Depth estimation is a computer vision technique used to infer three-dimensional structural information of a scene from two-dimensional images or videos. Its goal is to predict the depth value corresponding to each pixel in an image, i.e., the distance between the object's surface and the camera. Depth information helps machines understand spatial relationships within a scene and is a crucial foundational technology in fields such as 3D reconstruction, robot perception, autonomous driving, and augmented reality.

The development of depth estimation has evolved from traditional geometric methods to deep learning methods. Early methods mainly relied on sensors such as stereo vision, structured light, and LiDAR to obtain depth information through disparity calculation or active ranging, but these usually required additional hardware support. With the development of deep learning, researchers began to explore monocular depth estimation methods that directly predict depth using a single image. In 2014, researchers David Eigen et al. from New York University published a paper...Depth Map Prediction from a Single Image using a Multi-Scale Deep Network This paper proposes a monocular depth prediction method based on multi-scale convolutional neural networks. By combining global scene information with local detail features, it can predict depth maps from a single image, providing an important foundation for subsequent research on depth estimation based on deep learning.

Depth estimation techniques primarily address the lack of spatial scale in two-dimensional visual information. Depending on the input method, this task can generally be categorized into monocular depth estimation, binocular depth estimation, and multi-view depth estimation. Monocular methods rely solely on ordinary image input, inferring depth by learning visual cues such as object size, texture, occlusion relationships, and scene structure. Binocular methods, on the other hand, utilize the parallax relationships between multiple cameras to calculate spatial distance.

In recent years, thanks to the development of large-scale mixed dataset training and visual foundation models, the zero-shot generalization ability of monocular depth estimation models has been significantly improved—even when faced with scenes never encountered during training, they can provide relatively reliable depth predictions. Related technologies have been widely applied in scenarios such as autonomous driving environmental perception, robot navigation, 3D reconstruction, virtual reality (VR), and augmented reality (AR), providing crucial support for machines to acquire and understand the 3D world.