Pointline-indoor-layout
room structure and camera tilt detector
2 min readupdated
The pointline-indoor-layout model predicts the indoor layout from a single RGB photo, detecting floor and ceiling boundaries, vertical wall junctions, and room corners. It finds both visible elements and areas hidden behind furniture while completely ignoring the objects themselves. An additional TiltNet module estimates camera orientation (tilt and height) to reconstruct 3D spatial geometry. The model relies on a ConvNeXt-Tiny encoder and was trained exclusively on synthetic data (BlenderProc).
Key features
- Detection of hidden room lines (walls, floor, ceiling, and corners) even behind furniture
- Estimation of camera height and tilt angle (TiltNet) for 3D reconstruction
- Use of the ConvNeXt-Tiny architecture and a U-Net-like decoder
- Lightweight model weights (~33 MB for LayoutNet and ~40 MB for TiltNet)
What you can do
- Run inference on a room image to generate structure heatmaps
- Integrate the model into robotics for indoor geometry estimation via camera
- Calculate camera height and tilt angles for computer vision tasks
Example
import torch
from model import load_layout, load_tilt
layout = load_layout("layout.safetensors", "cuda")
prob = torch.sigmoid(layout(x)) # x: (B,3,480,640) RGB -> (B,8,240,320)
Read more
DetailsBenchmarks: F1 for visible structures 0.97–0.99, for full structures (including hidden ones) 0.92–0.96.
TiltNet error: median angle error 0.50°, median relative camera height error 6.6%.
Limitations: trained only on synthetic data, qualitative evaluation on real photos performed visually only.
License: Apache-2.0.










Comments