Research map / Image generation / Image restoration
Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Generates a 3,200-pair underwater depth dataset (Atlantis) by conditioning Stable Diffusion v1.5 with a Depth2Underwater ControlNet on terrestrial depth maps, then trains depth models on it to improve underwater depth estimation.
Problem
Monocular depth estimation performs poorly on underwater scenes due to scarce paired underwater image-depth data; existing synthetic datasets (style transfer) have large domain gaps and real datasets are small and costly.
Approach
Use Stable Diffusion with a ControlNet trained on {underwater image, MiDaS depth, BLIP2 caption} triplets to generate photorealistic non-existent underwater scenes that follow the layout of terrestrial depth maps, producing paired underwater image and accurate depth data for supervised training.
Key results
- Models trained on Atlantis outperform terrestrial-pretrained counterparts on Sea-thru and SQUID across all four depth models; e.g., iDisc RMSE on Sea-thru drops from 5.891 (KITTI) and 3.144 (NYU) to 1.371.
- LLaVA v1.5 judged 100% of Atlantis images as underwater versus 18% of UW-GAN images.
- Depth from Atlantis-trained models improves Sea-thru-based underwater image enhancement over KITTI-based depth, which produced reddish colors and artifacts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train depth models from scratch on Atlantis instead of terrestrial pretraining | iDisc, NeWCRFs, IEBins, VA-DepthNet · Sea-thru D3/D5 and SQUID | RMSE | Helped |
| train depth models on Atlantis instead of KITTI pretraining | VA-DepthNet · SQUID | RMSE: 8.753 to 2.666 | Helped |
| use Depth2Underwater ControlNet generated data (Atlantis) for depth training with Sea-thru evaluation | SIlog metric · Sea-thru D3/D5 | SIlog | Mixed |
| generated underwater images by SD+ControlNet instead of GAN-based style transfer | domain gap measured by LLaVA v1.5 · 14490 images, UW-GAN vs Atlantis | fraction judged underwater by LLVM: 100% vs 18% | Helped |
Surprising findings
- Atlantis is small (3,200 pairs) yet trained depth models beat models pretrained on large terrestrial datasets on underwater test sets.
- Generated scenes are non-existent underwater scenes built from terrestrial depth layouts rather than style-transferred terrestrial images.
Limitations
- Atlantis is small in scale and diversity compared to terrestrial datasets, and the authors suggest expansion.
- Depth labels come from MiDaS inverse relative depth capped at 20 m, so depth is pseudo-labeled rather than metric ground truth.
- Evaluation of domain gap relies on a single LLVM judge (LLaVA v1.5) rather than human assessment.
Experimental setup
Datasets: UIEB DIODE-outdoor KITTI NYU Depthv2 Sea-thru SQUID UW-GAN generated images
Benchmarks: Sea-thru D3 and D5 SQUID UIEB test set (qualitative) KITTI NYU Depthv2
Metrics: RMSE RMSElog A.Rel S.Rel log10 SIlog delta1 delta2 delta3 LLaVA yes-rate for underwater domain
Baselines: iDisc NeWCRFs IEBins VA-DepthNet UW-GAN MiDaS
Builds on: Stable Diffusion ControlNet MiDaS BLIP-2 Sea-thru
Closest papers by meaning
- Sea-ing in Low-light (CVPR 2025 · similarity 0.88)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.87)
- DepthFM: Fast Generative Monocular Depth Estimation with Flow Matching (AAAI 2025 · similarity 0.86)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.86)
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · similarity 0.86)
- Plenodium: Underwater 3D Scene Reconstruction with Plenoptic Medium Representation (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.