paper-conference

GVDepth leverages novel canonical representation to disentangle depth from camera parameters, ensuring consistency across diverse camera setups. Depth is estimated via probabilistic fusion of intermediate representations stemming from object size and vertical position cues, achieving accurate and generalizable predictions across multiple datasets and camera configurations. Notably, GVDepth achieves accuracy comparable to SotA zero-shot methods, while training with a single dataset collected with a single camera setup.

Self-attention-based approach for Monocular Depth Estimation integrates novel camera parameter embeddings to improve accuracy and generalization across varying vehicle-camera setups, supported by a new dataset from the CARLA simulator with diverse camera configurations.

A lightweight monocular odometry method that combines deep depth predictions with a feature-based geometrical framework, achieving metrically scaled trajectories, state-of-the-art accuracy, and robust generalization across different camera setups, as demonstrated on KITTI and KITTI-360 datasets.