3D reconstruction based on hierarchical reinforcement learning with transferability

3D reconstruction is extremely important in CAD (computer-aided design)/CAE Engineering)/CAM manufacturing). For interpretability, reinforcement learning (RL) used to reconstruct shapes from images by a series of editing actions. However, typical applications RL for face problems. The search space will increase exponentially with the action due curse dimensionality, which leads low performance, especially complex spaces reconstruction. Additionally, most works involve training specific agent each shape class without related experiences others. Therefore, we present hierarchical approach transferability (HRLT3D). First, actions are grouped into macro that can be chosen top-agent. Second, task accordingly decomposed hierarchically simplified sub-tasks solved sub-agents. Different classical (HRL), propose sub-agent based on augmented state (ASS-Sub-Agent) replace set sub-agents, speed up process shared and having fewer parameters. Furthermore, ASS-Sub-Agent more easily transferred data other classes diverse states tasks. experimental results public dataset show proposed HRLT3D performs overwhelmingly better than recent baselines. More impressingly, experiments also demonstrate extreme our among different classes.