Sparse Neural Generative Inference Based Pose Estimation

Current state-of-the-art pose estimation techniques often rely on end-to-end trained neural network architectures to infer a 6-DOF pose. These architectures, while effective, are computationally expensive and often require a large number of parameters. Instead, we propose formulating the pose estimation problem as a differentiable particle filter. This approach learns a per-particle latent embedding which infers its respective pose, object likelihood, and re-sampling objective iteratively. Our proposed method should decrease the tight coupling between input size and model architecture, and also improve the training and inference time scalability of the network. Furthermore, the proposed approach allows for reasoning across a distribution of inferred poses represented by the different particles.