yxma/gelsight-mini-pretrain

GelSight Mini Pretrain 是一个触觉感知数据集，包含约853,000个来自GelSight Mini传感器的触觉RGB帧，汇集了12个公共来源的数据，并使用统一的Parquet模式组织。该数据集专为自监督表示学习（如VAE、MAE、SimCLR、DINO）设计，每个帧都经过接触过滤、通道归一化，并以JPEG q92重新编码。数据来源包括真实数据（如FoTA、3DCal、FEATS、GelSLAM等）和模拟数据（如sim_tactile_mnist、sim_starstruck），覆盖标记和无标记凝胶类型，包含姿态、力、对象名称等可选标签。数据集支持多种应用，如自监督预训练、姿态/力回归、模拟到真实迁移以及标记不变性研究。GelSight Mini Pretrain is a tactile sensing dataset comprising ~853K GelSight Mini tactile RGB frames from 12 public sources, unified under one parquet schema. Built for self-supervised representation learning (e.g., VAE, MAE, SimCLR, DINO), every frame is contact-filtered, channel-normalized, and re-encoded as JPEG q92. It includes both real data (e.g., FoTA, 3DCal, FEATS, GelSLAM) and simulated data (e.g., sim_tactile_mnist, sim_starstruck), covering markered and markerless gel variants, with optional labels such as pose, force, and object names. The dataset supports various applications, including self-supervised pretraining, pose/force regression, sim-to-real transfer, and marker-invariance studies.