Source-linked AI summary
CubiCasa5K: A Dataset and an Improved Multi-Task Model for Floorplan Image Analysis
Ahti Kalervo, Juha Ylioinas, Markus Häikiö, Antti Karhu, Juho Kannala
TL;DR
Floorplan parsing needs representative large-scale datasets because rasterization removes geometric and semantic structure, while existing annotated datasets are small. This paper introduces CubiCasa5K and an improved multi-task CNN, reporting a dataset over 5× larger than existing alternatives and state-of-the-art performance.
Problem
Existing floorplan datasets are small and limited in object diversity and annotation quality, hindering investigation of automatic floorplan image analysis.
Method
The paper constructs CubiCasa5K with dense polygon annotations and uses an encoder-decoder multi-task CNN with automatically weighted task losses.
Results
CubiCasa5K is over 5× larger and covers over 80 object categories, while the improved multi-task CNN yields state-of-the-art performance.
Takeaways & Limitations
The released dataset and implementations provide richer tools for investigating automatic floorplan image analysis.
Abstract
from arXiv · showhide
Better understanding and modelling of building interiors and the emergence of more impressive AR/VR technology has brought up the need for automatic parsing of floorplan images. However, there is a clear lack of representative datasets to investigate the problem further. To address this shortcoming, this paper presents a novel image dataset called CubiCasa5K, a large-scale floorplan image dataset containing 5000 samples annotated into over 80 floorplan object categories. The dataset annotations are performed in a dense and versatile manner by using polygons for separating the different objects. Diverging from the classical approaches based on strong heuristics and low-level pixel operations, we present a method relying on an improved multi-task convolutional neural network. By releasing the novel dataset and our implementations, this study significantly boosts the research on automatic floorplan image analysis as it provides a richer set of tools for investigating the problem in a more comprehensive manner.
1 Introduction
Automatic floorplan analysis aims to recover structural and semantic information lost when vector drawings are rasterized. The paper addresses limited training data by introducing a larger, richly annotated dataset and an improved multi-task CNN.
- Rasterizing floorplans discards structured geometric and semantic information needed for applications such as virtual tours and 3D model creation.
- Existing floorplan analysis models use deep CNNs and multi-task heads, but available datasets contain fewer than 1000 images.
- CubiCasa5K contains 5000 floorplan images with polygon annotations covering over 80 object classes across three image categories.
- The dataset is over five times larger than the previously largest dataset and varies more in apartment types and drawing styles.
- The proposed method combines an encoder-decoder CNN with multi-task uncertainty loss to automatically determine task-loss weights.
2 Related Work
Related work shifted floorplan parsing from hand-crafted image processing toward CNN-based prediction. The paper focuses on improving network architecture and multi-task objectives while addressing the field’s limited and narrow datasets.
- CNN-based floorplan parsing replaced low-level image processing and heuristics with learned labeling and localization.
- Prior work also combined segmentation, object detection, and character recognition, either in isolated networks or shared multi-task architectures.
- Floorplan parsing improvements involve both network architecture and training objective, including U-net, residual, and hourglass designs.
- Multi-task learning can improve training, but performance depends strongly on the relative weighting of each task’s loss.
- Existing annotated datasets contain fewer than 1K samples and have limited object diversity, annotation consistency, and accuracy.
3 CubiCasa5K: A Novel Floorplan Dataset
CubiCasa5K is a 5,000-image floorplan dataset with rich vector annotations, broad class coverage, and quality-control procedures, paired with an improved multi-task parsing pipeline.
- Dataset composition: 5,000 floorplans are divided into high quality architectural, high quality, and colorful categories, with 4,200 training, 400 validation, and 400 test images.Annotations are provided as per-image SVG vector graphics containing semantic and geometric information.
- Annotations: Each sample undergoes two-stage quality assurance covering annotation placement accuracy and label correctness.The second review is performed by a different QA person, who corrects errors remaining after the annotator’s review.
- Dataset statistics: CubiCasa5K is reported as the largest and most versatile publicly available annotated floorplan dataset, with statistics covering class distributions, image resolutions, and sample complexity.Figures compare room, icon, wall, and resolution distributions with the dataset of Liu et al.
- Parsing pipeline: The parsing pipeline combines an hourglass encoder-decoder CNN with multi-task outputs and trainable uncertainty-based loss weighting.Predictions are post-processed into walls, rooms, icons, doors, and windows with locations, dimensions, and category labels.
5 Results
The experiments evaluate the method on the prior benchmark and CubiCasa5K, reporting improved benchmark performance and benefits from test-time augmentation. CubiCasa5K results compare raw and polygonized segmentations using accuracy and IoU metrics.
- The model significantly improves the results reported on the prior benchmark dataset.
- Test-time augmentation averages four predictions from 90-degree image rotations and appears beneficial with and without integer programming.
- Corrected annotations reveal that the prior method performed better than originally reported, while the authors’ model remained clearly better in comparison.
- CubiCasa5K Experiments: CubiCasa5K experiments use coupled targets covering 12 room classes and 11 icon classes, with the same heatmap regression layers and training objective as the previous experiment.
- Evaluation metrics: Evaluation reports overall accuracy, mean accuracy, and mean IoU for both raw segmentations and polygonized instances.
- CubiCasa5K results: Raw segmentation test scores are clearly better than polygonized segmentation-instance scores because missed or misplaced junctions prevent polygon creation.
6 Conclusions
The paper introduces CubiCasa5K, a larger and more varied annotated floorplan dataset, together with an improved multi-task convolutional neural network and baseline results. Future work includes adding object detection and directly inferring floorplan elements as polygons.
- CubiCasa5K is over 5× larger than existing annotated floorplan datasets and covers over 80 object categories with more varied annotations.
- The paper provides baseline results from an improved multi-task convolutional neural network yielding state-of-the-art performance.
- Future directions: Future work proposes integrating an object detector and directly inferring floorplan elements as polygons.