Overview
Furniture Recommendation from Space was an SKT FLY AI prototype for recommending furniture that matches a user’s room. The work combined room-image analysis, furniture/object detection, style inference, and single-view 3D reconstruction experiments for furniture product images.
Goal
- Detect furniture, objects, and spatial regions from user-captured room images.
- Infer the room’s interior style from the detected elements.
- Recommend furniture products that fit the inferred atmosphere.
Role
- Implemented the SKT FLY AI prototype flow for room-aware furniture recommendation.
- Developed furniture-detection models with MMDetection.
- Built an AI-serving pipeline with BentoML.
- Ran single-view 3D reconstruction experiments on crawled furniture product images.
- Connected detection, style classification, reconstruction outputs, recommendation, and serving into one prototype flow.
3D reconstruction idea
The reconstruction task starts from a difficult setting: only a single product image is available, without known camera pose or multi-view observations. The project therefore used category-level priors and pose candidates to approximate a 3D shape that visually matches the input image.
A candidate set of pose and shape parameters was evaluated against the single-view image, selecting the configuration that best matched the observed furniture appearance.
Materials
The project visuals show the service UI, recommendation flow, detection process, and single-view 3D reconstruction experiments that supported the room-aware furniture recommendation prototype.
