AutoSweep: Recovering 3D Editable Objects from a Single Photograph.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Modeling

Background:

• Existing methods for 3D object recovery from images often produce depth maps, point clouds, or mesh surfaces, lacking semantic part information and editability.
• Human-made objects are frequently composed of distinct parts that can be represented by generalized primitive shapes.

Purpose of the Study:

• To develop a fully automatic framework for extracting editable 3D objects with semantic parts from a single photograph.
• To recover 3D models of primitive-shaped objects, specifically generalized cuboids and generalized cylinders.

Main Methods:

• A novel instance-aware segmentation network (GeoNet) was developed for accurate separation of object parts into profiles and bodies.
• A key stage involves simultaneously identifying profile-body relationships and reconstructing 3D parts by sweeping profiles along body contours.
• Joint optimization of geometry is performed to align with the recovered masks.

Main Results:

• The GeoNet successfully outputs smooth, part-level masks, enabling accurate instance segmentation.
• The proposed method effectively recovers 3D parts by combining profile sweeping and mask-aligned geometry optimization.
• Qualitative and quantitative experiments demonstrate high-quality 3D model recovery.

Conclusions:

• The framework provides a significant advancement in recovering editable 3D objects with semantic parts from single images.
• The approach outperforms existing methods in both instance segmentation and 3D reconstruction accuracy.
• This work enables direct editing of recovered 3D models, opening new possibilities in computer graphics and vision applications.