Sunday 02 March 2025
The ability to estimate depth from a single image has long been a challenge in computer vision, and recent advancements have brought us closer than ever to achieving this feat. A new framework, called Depth Any Camera (DAC), has shown remarkable success in estimating metric depth from any type of camera, including those with large fields of view.
The issue with previous methods is that they relied on specialized training data for specific types of cameras, making them ineffective when applied to other types of cameras. DAC addresses this limitation by using a unified image representation called Equi-Rectangular Projection (ERP), which allows the model to be trained on perspective images and then generalized to fisheye and 360-degree cameras without requiring additional training data.
The ERP conversion is the key innovation behind DAC, allowing for efficient online augmentation in ERP space. This process involves converting the input image into a rectangular patch that can be processed by the network, while also preserving the original image’s distortion. The resulting ERP depth maps can then be up-projected into 3D point clouds using each ERP coordinate’s ray direction.
To evaluate the performance of DAC, researchers tested it on several large-Field-of-View (FoV) datasets, including Matterport3D, Pano3D-GV2, Scannet++, and KITTI360. The results were impressive, with DAC achieving state-of-the-art zero-shot metric depth estimation accuracy on these datasets.
In addition to its impressive performance, DAC also offers several practical benefits. For example, it can be used for visualization purposes, such as mapping ERP patches back to the original fisheye space. This allows for more accurate and visually consistent scale in depth visualization results.
DAC’s ability to generalize across different camera types has significant implications for a range of applications, from autonomous driving to augmented reality. By allowing developers to estimate metric depth from any type of camera, DAC opens up new possibilities for 3D reconstruction and object detection.
One potential application of DAC is in the field of robotics, where accurate depth estimation can be crucial for navigation and obstacle avoidance. With DAC, robots could potentially use a single camera to navigate complex environments with ease, without requiring specialized cameras or additional sensors.
In summary, Depth Any Camera (DAC) represents a significant step forward in computer vision, enabling the estimation of metric depth from any type of camera.
Cite this article: “Depth Any Camera: Estimating Metric Depth from Any Type of Camera”, The Science Archive, 2025.
Computer Vision, Depth Estimation, Camera Types, Erp, Perspective Images, Fisheye Cameras, 360-Degree Cameras, Metric Depth, Autonomous Driving, Augmented Reality
