PixVOD: Pixel-Distributed Direct Visual Odometry and Depth Estimation
Researchers have developed PixVOD, a novel approach to visual odometry and depth estimation that distributes computations directly within each pixel of a sensor. This method leverages Gaussian Belief Propagation (GBP) for parallel processing, enabling pixels to exchange information and reach consensus on camera motion and depth. A keyframe-like anchoring mechanism is introduced to ensure geometric stability during optimization, allowing for consistent updates. The system has been evaluated on realistic datasets, showing its potential for on-sensor processing. AI
IMPACT Introduces a novel on-sensor processing approach for visual odometry, potentially reducing data transmission needs and enabling real-time depth estimation.