| Title |
Generative data augmentation for dual-plane monocular depth estimation of non-Lambertian surfaces |
| Authors |
Bankauskas, Vilius ; Kriščiūnas, Andrius |
| Full Text |
|
| Is Part of |
CEUR workshop proceedings: IVUS 2026: proceedings of the 31st international conference on information society and university studies (IVUS 2026), Kaunas, Lithuania, May 12, 2026 / edited by Audrius Lopata,Tomas Krilavičius, Ilona Veitaitė.. Aachen : CEUR-WS. 2026, vol. 4317, p. 1-10.. ISSN 1613-0073 |
| Keywords [eng] |
Monocular Depth Estimation ; Generative AI ; Data Augmentation ; Non-Lambertian Surfaces ; Synthetic Data ; Diffusion Models ; Depth Mapping |
| Abstract [eng] |
leading to critical failures in depth maps. Traditional data acquisition methods, including LiDAR, fail to capture accurate ground truth for these objects due to physical light-bending properties. To address this lack of data, a generative data augmentation pipeline is proposed, utilizing advanced diffusion models to synthesize a dual-plane dataset from existing high-quality RGB-D s cans. B y e ither i npainting new transparent objects (forcing the model to estimate the background depth) or performing a material transfer on existing objects (providing ground truth for the transparent surface depth), two distinct training datasets are created. The proposed methodology uses the latest generative architectures for their superior curvature preservation and inpainting accuracy. Evaluated using the high-resolution DIODE indoors dataset, this generative augmentation offers an optimized, scalable solution to overcome the lack of data caused by sensor limitations, enabling training of MDE models capable of accurately perceiving complex scenes with transparent elements. |
| Published |
Aachen : CEUR-WS |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
