SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema

Key points

• RGB-based category-level 9D pose estimation remains highly challenging because a monocular system must generalize across unseen object instances while jointly inferring 3D translation, 3D rotation, and 3D size.1
• Unlike instance-level pose estimation, the category-level setting requires the network to learn both what is shared across objects within a category and how individual instances vary in structure.1
• This challenge is amplified in the RGB-only setting, where pose cues, category-level commonality, and instance-specific variation are entangled in appearance without direct geometric constraints, making it difficult for existing frameworks to learn a representation that is truly aligned with the category-level estimation objective.1
• SchemaPose equips the estimation pipeline with an explicit parametric schema for modeling category-level regularities together with structured instance variation, and uses the inferred schema variables to guide pose-related prediction in an end-to-end query-based framework.1

Sources (1)

• [1]SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 01:29 PM

  To address this problem, we propose SchemaPose, a schema-guided framework for RGB-based category-level pose estimation.

  RGB-based category-level 9D pose estimation remains highly challenging because a monocular system must generalize across unseen object instances while jointly inferring 3D translation, 3D rotation, and 3D size.

Extractive summary: sentences quoted from the sources.