Global Completion of Occupancy Maps Using Generative Models
ELECTRONIC LETTERS ON COMPUTER VISION AND IMAGE ANALYSIS (ELCVIA), vol.24, no.2, pp.273-285, 2026 (Scopus)
- Publication Type: Article / Article
- Volume: 24 Issue: 2
- Publication Date: 2026
- Doi Number: 10.5565/rev/elcvia.2162
- Journal Name: ELECTRONIC LETTERS ON COMPUTER VISION AND IMAGE ANALYSIS (ELCVIA)
- Journal Indexes: Scopus, Directory of Open Access Journals, DIALNET
- Page Numbers: pp.273-285
- Open Archive Collection: AVESIS Open Access Collection
- Yıldız Technical University Affiliated: Yes
Abstract
We address global occupancy–map completion from partial, sequential observations by framing it asconditional image generation. Using HouseExpo/SUNCG floorplans, we build a Gazebo+ROS pipeline thatcollects time–ordered exploration data and release a 128×128 dataset for reproducible benchmarking.Six model families are compared under a unified setup: Conditional VAE (CVAE), pix2pix–hingeConditional GAN (CGAN), RePaint–style diffusion (DDPM), Vision Transformer (ViT) encoder–decoder,residual UNet, and a geometric, training–free baseline (GEOM). UNet achieves the lowest pixelwise errors(L1/MSE/RMSE) and the best discrete scores (IoU/F1/Accuracy). GEOM, despite being training-free, at-tains the highest SSIM—demonstrating that explicit geometric priors can rival data-driven learning whenstructural regularity dominates. Diffusion offers strong perceptual quality but at higher computational cost.Our findings reveal that occupancy-map completion is fundamentally a structure-exploitation task: con-volutional architectures inherently favor rectilinear layouts, while attention-based models require more datato learn comparable priors. We release the dataset, ROS pipeline, and all implementations to enable repro-ducible research in generative mapping.