Orient Anything V2: Unifying Orientation and Rotation Understanding

TL;DR

Orient Anything V2 advances unified object orientation and rotation understanding by coupling a scalable synthetic data engine with a symmetry-aware learning objective and a multi-frame architecture for relative rotation estimation. The data engine produces 600k diverse assets with robust 0-to-N front-face annotations, while the framework directly models rotational symmetry and predicts relative rotations across frames. Across 11 benchmarks, the model achieves state-of-the-art zero-shot performance in orientation, 6DoF pose estimation, and symmetry recognition, demonstrating strong generalization to open-world objects and challenging viewpoints. This work broadens the applicability of orientation estimation to downstream tasks such as robotics, AR/VR, and autonomous perception by enabling robust, symmetry-aware, and multi-view understanding from minimal input.

Abstract

This work presents Orient Anything V2, an enhanced foundation model for unified understanding of object 3D orientation and rotation from single or paired images. Building upon Orient Anything V1, which defines orientation via a single unique front face, V2 extends this capability to handle objects with diverse rotational symmetries and directly estimate relative rotations. These improvements are enabled by four key innovations: 1) Scalable 3D assets synthesized by generative models, ensuring broad category coverage and balanced data distribution; 2) An efficient, model-in-the-loop annotation system that robustly identifies 0 to N valid front faces for each object; 3) A symmetry-aware, periodic distribution fitting objective that captures all plausible front-facing orientations, effectively modeling object rotational symmetry; 4) A multi-frame architecture that directly predicts relative object rotations. Extensive experiments show that Orient Anything V2 achieves state-of-the-art zero-shot performance on orientation estimation, 6DoF pose estimation, and object symmetry recognition across 11 widely used benchmarks. The model demonstrates strong generalization, significantly broadening the applicability of orientation estimation in diverse downstream tasks.

Figures (13)

• Figure 1: Overview of Orient Anything V2. We upgrade the foundation orientation estimation model from both Data and Model perspectives. It unifies the understanding of object orientation and rotation, achieving better estimation accuracy and gaining the New Features to handle rotational symmetry and relative rotation. Zoom in for the best view.
• Figure 2: Real assets from Objavese suffer from (a) low-quality texture and (b) limited realism.
• Figure 3: Overview of 3D Asset Synthesis Pipeline. We begin with class tags and use a series of advanced generative models to progressively generate high-quality 3D assets.
• Figure 4: Overview of Robust Annotation Pipeline. "Pseudo Label" visualizes the azimuth direction of pseudo labels and objects in the horizontal plane. By fitting the pseudo labels to standard periodic distribution, we can robustly derive the orientation and symmetry label. Human calibration is only required for categories with symmetry inconsistencies.
• Figure 5: Framework of Orient Anything V2. One or two input frames are tokenized by DINOv2 and then jointly encoded using transformer blocks. We finally employ MLP heads to predict the orientation or rotation distributions from the encoded learnable tokens of each frame.