Learning and Optimizing Camera Pose

Abstract: Plenty of computer vision applications involve assessing the position and orientation, i.e. the pose , of one or several cameras, including object pose estimation, visual  localization, and structure-from-motion. Traditionally, such problems have often been addressed by detection, extraction, and matching of image keypoints, using handcrafted local image features such as the scale-invariant feature transform (SIFT), followed by robust fitting and / or optimization to determine the unknown camera pose(s). Learning-based models have the advantage that the they can learn from data what cues or patterns are relevant for the task, beyond the imagination of the engineer. However, compared with 2D vision tasks such as image classification and object detection, applying machine learning models to 3D vision tasks such as pose estimation has proven to be more challenging. In this thesis, I explore pose estimation methods based on machine learning and optimization, from the aspects of quality, robustness, and efficiency. First, an efficient and powerful graph attention network model for learning structure-from-motion is presented, taking image point tracks as input. Generalization capabilities to novel scenes is then demonstrated, without costly fine-tuning of network parameters. Combined with bundle adjustment, accurate reconstructions are acquired, significantly faster than off-the-shelf incremental structure-from-motion pipelines. Second, techniques are presented for improving the equivariance properties of convolutional neural network models carrying out pose estimation, either by intentionally applying radial distortion to images to reduce perspective effects, or via a geometrically sound data augmentation scheme corresponding to camera motion. Next, the power and limitations of semidefinite relaxations of pose optimization problems are explored, notably leading to the conclusion that absolute camera pose estimation is not necessarily solvable using the considered semidefinite relaxations, since while they tend to almost always be tight in practice, counter-examples do indeed exist. Finally, a rendering-based object pose refinement method is presented, robust to partial occlusion due to its implicit nature, followed by a method for long-term visual localization, leveraging on a semantic segmentation model to increase the robustness by promoting semantic consistency of sampled point correspondences.