Deep Depth from Uncalibrated Small Motion Clip.

Area of Science:

• Computer Vision
• Machine Learning
• Photogrammetry

Background:

• Depth map estimation is crucial for 3D scene understanding.
• Existing methods often require pre-calibrated cameras and struggle with small viewpoint variations.
• Traditional approaches combine camera pose estimation and dense reconstruction.

Purpose of the Study:

• To propose a novel, self-calibrating approach for high-quality depth map inference.
• To enable accurate camera pose computation without prior camera calibration.
• To develop an effective deep learning model for dense depth reconstruction.

Main Methods:

• A self-calibrating bundle adjustment method for camera pose estimation with small motion.
• A convolutional neural network, DPSNet (Deep Plane Sweep Network), for dense depth reconstruction.
• DPSNet utilizes a plane sweep algorithm to build and regularize a cost volume from deep features for end-to-end depth regression.

Main Results:

• The proposed method achieves state-of-the-art performance on challenging datasets.
• The self-calibrating bundle adjustment successfully computes camera poses without calibration.
• DPSNet effectively integrates traditional multi-view stereo concepts into a deep learning framework.

Conclusions:

• The novel approach provides high-quality depth maps from images with small viewpoint variations.
• The method eliminates the need for camera calibration, simplifying the depth estimation pipeline.
• Combining geometric principles with deep learning yields superior results in depth estimation.