StereoAdapter: Adapting Stereo Depth Estimation to
Underwater Scenes

Zhengri Wu^1* Yiran Wang^2* Yu Wen^1* Zeyu Zhang^3*† Biao Wu¹ Hao Tang^3‡
¹ AI Geeks    ²Australian Centre for Robotics    ³Peking University
^* Equal contribution.    ^† Project lead.    ^‡ Corresponding author.

[Paper]      [Code]     [Dataset]     [Model]

TL;DR: StereoAdapter is a self-supervised adaptive model that allows robust underwater depth estimation.

Real World Results

We evaluate StereoAdapter on a BlueROV2 across three obstacle layouts and three motion trajectories.

BEV

Right Camera

Left Camera

Ground Truth

StereoAdapter

Layout #1

Layout #2

Layout #3

Method

Detailed architecture of the StereoAdapter: (a) two-stage self-supervised training pipeline; (b) update mechanism with LoRA.

Benchmark Results

Visualization results of Stereo Depth Estimation Methods on SQUID and TartanAir.

UW-StereoDepth-40K

Example from the high-quality UW-StereoDepth-40K dataset.

StereoAdapter: Adapting Stereo Depth Estimation to Underwater Scenes