Dataset Open Access

PolarDepth

Polarized Transparent Object Dataset

Wen Dong, Haiyang Mei, Yinglian Ji, Zijun Zhang, Wenyuan Zhang, Pengwei Luo, Bo Dong, Shengfeng He, Xin Yang

PolarDepth: Monocular Transparent Object Depth from Polar-Physics Priors

In Forty-third International Conference on Machine Learning (ICML 2026).

Paper

https://icml.cc/virtual/2026/poster/61742

Introduction

We construct the Polarized Transparent Object Dataset (PTOD), a novel pixel-aligned RGB-polarization dataset for transparent object perception. PTOD is captured using a polarizer-array camera that records RGB images along with DoLP and AoLP maps at a resolution of 612 × 512. An active stereo depth sensor provides a metric depth map.

PTOD contains 2,878 scenes, each featuring at least one transparent glass object with diverse shapes and placements. To overcome transparency-induced failures in depth sensing, we adopt a replacement-based acquisition strategy that enables accurate geometry capture of transparent objects.

High-quality ground-truth depth maps are obtained by refining predictions from Depth Anything v2, using sensor depth as a metrically accurate reference.

In addition to depth annotations, PTOD provides manually labeled segmentation masks for transparent regions, enabling joint learning of depth estimation and transparent object localization.


Please enter your application information for dataset