Syn-TODD

该数据集是一个大型的透明物体检测数据集，适用于训练具备RGB、RGB-D、立体和多视角模态的网络。它包含了113,772对立体图像，覆盖了1996个不同的场景。数据集的特色在于透明物体的程序化生成、场景复杂度的域随机化，以及丰富的标注信息，包括2D和3D边界框、物体姿态、实例分割、深度和表面法线。该数据集规模庞大，拥有113,772对立体图像和1996个不同的场景，任务涵盖了深度估计、分割、姿态估计以及3D物体检测。

This dataset is a large-scale transparent object detection dataset suitable for training networks with RGB, RGB-D, stereo, and multi-view modalities. It contains 113,772 stereo image pairs covering 1,996 distinct scenes. The notable features of this dataset include procedural generation of transparent objects, domain randomization of scene complexity, and rich annotation information including 2D and 3D bounding boxes, object poses, instance segmentation, depth maps, and surface normals. With a massive scale of 113,772 stereo image pairs and 1,996 distinct scenes, it supports tasks such as depth estimation, segmentation, pose estimation, and 3D object detection.

搜集汇总

背景与挑战

背景概述

Syn-TODD是一个用于透明物体检测的大规模数据集，支持RGB-D、立体和多视图RGB三种模式，适用于深度估计、分割和姿态估计等任务的网络训练。