Research map / 3D vision / 3D generation
Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries
Yuanwen Yue, Theodora Kontogianni, Konrad Schindler, Francis Engelmann
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Predict rooms as a set of ordered vertex sequences with two-level queries rather than single-level queries (ablation) | RoomFormer · Structured3D val | Room precision: 74.4 -> 96.5 | Improved |
| Use multi-scale feature maps instead of only the last-stage feature map (ablation) | RoomFormer · Structured3D val | Room precision: 93.9 -> 96.5 | Improved |
| Iteratively refine polygon coordinates and use updated positions as reference points (ablation) | RoomFormer · Structured3D val | Room precision: 94.8 -> 96.5 | Improved |
| Add coordinate regression loss alongside classification and rasterization losses (ablation) | RoomFormer · Structured3D val | Room precision: 84.3 -> 96.5 | Improved |
| Add polygon rasterization loss (ablation) | RoomFormer · Structured3D val | Room precision: 96.0 -> 96.5 | Improved |
| Allow self-attention between vertices across different polygons (ablation) | RoomFormer · Structured3D val | Room F1: 94.8 -> 95.9 | Improved |
| Increase room-level and corner-level query counts from M=15, N=30 to M=20, N=40 (ablation) | RoomFormer · Structured3D val | inference time: 8.4 -> 10.8 | Hurt |
| Use separate room and line decoders with two-level line queries (ablation) | RoomFormer semantic-rich extension · Structured3D test, semantic-rich | Room type F1: 70.7 -> 71.4 | Improved |
| Use single-level rather than two-level queries in the separate line decoder (ablation) | RoomFormer semantic-rich extension · Structured3D test, semantic-rich | Room type F1: 71.4 -> 74.4 | Improved |
| Train using tight room layouts rather than the original room layout representation (ablation) | RoomFormer · Structured3D test | Angle F1: 81.2 -> 84.2 | Improved |
| Train using tight room layouts rather than the original room layout representation (ablation) | RoomFormer · Structured3D test | Room F1: 97.3 -> 96.7 | Hurt |
| Replace the original RoomFormer with holistic polygon prediction instead of a corner-and-edge pipeline | RoomFormer versus HEAT · SceneCAD val, trained on Structured3D | Room IoU: 52.5 -> 74.0 | Improved |
Closest papers by meaning
- PolyRoom: Room-aware Transformer for Floorplan Reconstruction (ECCV 2024 · similarity 0.93)
- FRI-Net: Floorplan Reconstruction via Room-wise Implicit Representation (ECCV 2024 · similarity 0.91)
- CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction (NeurIPS 2025 · similarity 0.90)
- SLIBO-Net: Floorplan Reconstruction via Slicing Box Representation with Local Geometry Regularization (NeurIPS 2023 · similarity 0.90)
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.88)
- GSDiff: Synthesizing Vector Floorplans via Geometry-enhanced Structural Graph Generation (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
