Segmentation-Enhanced Depth Estimation Using Camera Model Based Self-supervised Contrastive Learning
Keywords: Contrastive Learning; Depth Estimation
Abstract: Depth estimation is a key topic in the field of computer vision. Self-supervised monocular depth estimation offers a powerful method to extract 3D scene information from a single camera image, allowing training on arbitrary image sequences without the need for depth labels. However, monocular unsupervised depth estimation still cannot address the issue of scale and often requires ground truth for calibration.
In the deep learning era, existing methods primarily rely on relationships between images to train unsupervised neural networks, often overlooking the fundamental information provided by the camera itself. In fact, the intrinsic and extrinsic parameters of the camera can be used to compute depth information for the ground and its related areas based on physical principles. This information can offer rich supervisory signals at no additional cost. Additionally, by assuming that objects like people, cars, and buildings share the same depth as the corresponding ground, the physical depth of the entire scene can be inferred, and gaps in the depth map can be filled.
Since some areas may have depth estimation errors, to make full use of these regions, we introduce a contrastive learning self-supervised framework. This framework consists of two networks with the same structure: the Anchor network and the Target network. While calculating depth, the network also outputs semantic segmentation results to assist in computing the physics depth, which is then used as the label for the model. Semantic segmentation can identify dynamic objects, reducing photometric reprojection errors caused by moving objects. The predictions from the Anchor network are used as pseudo-labels for training the Target network. Reliability is determined by entropy, dividing the predicted depth into positive and negative samples to maximize the use of physics depth information.
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5866
Loading