GRIHA: Synthesizing 2-Dimensional Building Layouts from Images Captured using a Smart Phone

The paper proposes GRIHA, a fast and efficient framework that utilizes SLAM technology embedded in standard mobile ARCore libraries to generate 2D building layouts from simple RGB smartphone images, overcoming the hardware and occlusion limitations of existing methods.

The Gist

Imagine you are a detective trying to solve a mystery: "What does this room actually look like on paper?"

Usually, to draw a floor plan (a 2D map of a room), you have to do one of two tedious things:

1. The Old School Way: Walk around with a tape measure, write down every inch, and then spend hours drawing it on a computer.
2. The High-Tech Way: Use a special, expensive camera that can "see" depth (like a robot's eyes) or take a perfect, 360-degree panoramic photo where nothing blocks the view.

The Problem: Most of us don't have robot eyes or expensive cameras. And if you try to take a 360-degree photo in a messy room full of furniture, the sofa blocks the wall, the table hides the corner, and the photo becomes useless.

The Solution: Enter GRIHA (pronounced like "Gree-ha," meaning "Home" in Sanskrit). It's a new app that lets you create a perfect floor plan using just a regular smartphone and a few quick photos.

Here is how GRIHA works, broken down into simple analogies:

1. The "Magic Glasses" (ARCore & SLAM)

Imagine you are wearing a pair of magic glasses that know exactly where you are standing in the room, even if you are just walking around.

• How it works: GRIHA uses a technology built into modern Android phones called ARCore. It's like a GPS for inside a building. As you walk and take photos, the phone tracks your movement (how far you walked, how much you turned) using its internal sensors. It builds a "mental map" of your path without needing any special hardware.

2. The "Imagination Engine" (Depth Estimation)

A normal phone camera takes a flat, 2D picture. It doesn't know how far away the wall is.

• The Analogy: Imagine looking at a painting of a mountain. A normal eye sees a flat image. But GRIHA has a super-powered AI brain that looks at the flat photo and says, "I know this looks flat, but based on the shadows and textures, I can imagine how deep this room is."
• The Tech: It uses a pre-trained AI (a neural network) to guess the "depth map" of the room. It turns a flat photo into a 3D mental model, estimating how far the furniture is from the wall.

3. The "Puzzle Solver" (3D Reconstruction)

Now, the app has a bunch of flat photos and a guess about how deep they are. It needs to build a 3D model.

• The Analogy: Think of it like building a 3D sculpture out of clay. You take a photo of a corner, and the app sculpts a little piece of 3D clay for that corner. You take a photo of the door, it sculpts the door. It does this for every photo you take.
• The Twist: Because you took photos from different angles, the app uses your walking path (from Step 1) to stitch these 3D clay pieces together into one big, floating 3D model of the room.

4. The "Straightener" (Regularization)

Real life is messy. Walls might look slightly crooked in a photo because of the camera angle. But in the real world, walls are usually straight and meet at perfect 90-degree angles (like a grid).

• The Analogy: Imagine you drew a room on a piece of paper, but your hand shook, so the walls are wobbly. GRIHA is like a magical ruler that comes in, grabs the wobbly lines, and snaps them perfectly straight. It forces the walls to be straight and the corners to be perfect right angles, assuming the room follows "Manhattan rules" (like a city grid).

5. The "Door Detective" (Door Placement)

Finally, the app looks at your photos to find the doors.

• The Analogy: It's like a game of "Where's Waldo?" but for doors. It scans the images, spots the door, and then calculates exactly where that door should go on your new, straightened floor plan. It places the door symbol in the exact right spot relative to the walls.

Why is this a Big Deal?

• No Special Gear: You don't need a $2,000 depth camera. Your old smartphone works.
• Handles Messy Rooms: If you have a couch blocking a corner, other apps fail because they can't "see" the wall. GRIHA takes a photo of the other side of the room, figures out the geometry, and fills in the missing piece. It's like solving a puzzle even when some pieces are hidden.
• Super Fast: Instead of walking around the whole room scanning everything (which takes 10 minutes), you just snap 4 photos (one from each corner) and the app does the rest in seconds.

The Result

The paper tested this on real offices and labs. They found that GRIHA was just as accurate as the expensive apps that require you to walk around and manually measure everything, but it was much easier to use and didn't get confused by furniture blocking the view.

In short: GRIHA turns your smartphone into a magic architect that can draw your house's blueprint just by looking at a few snapshots.

Technical Summary

1. Problem Statement

The generation of 2D/3D Building Information Models (BIM) and floor plans from indoor scenes is a critical task for real estate, navigation, and augmented reality. However, existing solutions face significant limitations:

• Hardware Dependency: Most high-accuracy methods rely on RGB-D cameras (e.g., LiDAR, Kinect) or specialized depth sensors, which are not available on standard consumer smartphones.
• Occlusion Sensitivity: Methods relying on panoramic or 360° images often fail in cluttered environments (offices, homes) where furniture and decor occlude corners and walls, making it impossible to capture a complete, unobstructed view.
• User Burden: Current mobile apps often require continuous scanning, manual corner addition, and complex calibration, leading to user fatigue and potential measurement errors.

The Goal: Develop a system that can generate accurate 2D floor plans using only standard RGB images captured by a typical smartphone camera, without requiring depth hardware, panoramic modes, or extensive manual intervention, even in occluded environments.

2. Methodology: The GRIHA Framework

The proposed framework, GRIHA (Generating Room Interior of a House using ARCore), operates on the assumption of a "Manhattan World" (rooms with orthogonal walls). The pipeline consists of the following stages:

A. Data Acquisition & Pose Estimation

• Input: Users capture a few standard 2D photos of room corners and doors using a custom Android application built in Unity3D.
• Pose Tracking: The system leverages Google ARCore, which utilizes the phone's IMU and visual feature tracking (SLAM) to estimate the 6-DoF (Degrees of Freedom) camera pose (position and orientation) for each captured image in real-time.

B. Depth Estimation

• Since the input is monocular RGB, depth must be inferred.
• The system adapts a pre-trained Encoder-Decoder architecture (based on DenseNet-169) to predict depth maps from single RGB images.
• This allows the conversion of 2D pixels into 3D coordinates without physical depth sensors.

C. Edge and Boundary Detection

• To identify room geometry, the system uses LayoutNet (adapted from [31]), an encoder-decoder network.
• It predicts boundary maps (walls/edges) and corner maps.
• The network is enhanced with skip connections and additional predictors for room types to handle smaller fields of view and occluded corners.

D. 3D Reconstruction & Point Cloud Generation

• The predicted depth maps are fused with the RGB images and camera intrinsic parameters (focal length, center coordinates) to generate 3D point clouds for each partial scene.
• The formula used maps pixel coordinates $(u, v)$ and depth $Z$ to real-world coordinates $(X, Y, Z)$.

E. Regularization and Layout Synthesis

This is the core innovation for handling occlusion and noise:

1. Local Regularization: For each room's partial point cloud, the system clusters points (using K-means) to identify wall segments. It enforces Manhattan constraints, forcing lines to be orthogonal (90°) to correct estimation errors.
2. Global Transformation: Using the ARCore camera poses, all partial point clouds are transformed into a unified global coordinate system.
3. Global Regularization: The system constructs a global polygon and enforces perpendicularity between adjacent walls across the entire floor plan.
4. Boundary Alignment: Algorithms (Alg 2 & 3) identify points lying inside the global boundary and project them onto the boundary lines to ensure a clean, closed floor plan.

F. Door Detection

• To complete the floor plan, a YOLO object detection network (trained on the DoorDetect dataset) identifies doors in the images.
• The system calculates the relative position of the door on the wall using a scale-invariant ratio and maps this position onto the generated 2D floor plan.

3. Key Contributions

1. RGB-Only Floor Plan Generation: The first framework to generate 2D layouts from standard smartphone RGB images without requiring depth sensors or panoramic/360° capture modes.
2. Occlusion Robustness: By using multiple partial views rather than a single panoramic sweep, the system effectively handles cluttered rooms where furniture blocks corners, a common failure point for existing methods.
3. Minimal User Interaction: The system requires only 4 interactions (capturing 4 corners per room) and zero manual intervention (no manual corner drawing or height calibration), unlike competitors that require continuous scanning.
4. Cross-Platform Validation: The system was tested on multiple hardware platforms (Google Pixel 2 XL and Samsung A50), demonstrating platform independence.

4. Experimental Results

The authors evaluated GRIHA on three datasets: Classrooms, Labs (highly occluded), and Offices.

• Quantitative Accuracy:
  • Area Error: GRIHA achieved a Mean Absolute % Error (MAPE) of 3.12% for classrooms and 3.25% for offices, outperforming or matching commercial apps like Magic Plan, Tape Measure, and AR Plan 3D.
  • Aspect Ratio Error: GRIHA achieved 2.21% error for classrooms, significantly better than AR Plan 3D (5.20%).
  • Corner/Pixel Error: The system maintained low corner error (1.04% - 1.30%) and pixel error (3.38% - 4.15%).
• Comparative Performance:
  • While some commercial apps (like AR Plan 3D) performed slightly better in specific "Area" metrics for Labs (likely due to manual user correction), GRIHA provided superior results with no manual intervention.
  • Existing apps failed significantly in low-light or low-contrast environments where feature detection was impossible. GRIHA remained robust.
• Efficiency: The system is computationally efficient, leveraging built-in ARCore SLAM, and consumes battery power comparable to other mobile AR applications.

5. Significance and Conclusion

The GRIHA framework represents a significant step toward democratizing Building Information Modeling (BIM). By removing the dependency on expensive depth hardware and the strict requirement for occlusion-free panoramic views, it makes floor plan generation accessible to the average user with a standard smartphone.

• Practical Impact: It enables rapid, automated layout generation for real estate, interior design, and emergency response planning in cluttered, real-world environments.
• Future Work: The authors plan to relax the "Manhattan World" assumption to handle non-orthogonal (curved or angled) architectural designs and generalize the system for more complex scenes.

In summary, GRIHA successfully synthesizes 2D layouts from sparse, occluded RGB images by combining deep learning for depth/edge estimation with geometric regularization and AR-based pose tracking, offering a robust, user-friendly alternative to existing commercial solutions.