Multi-Object Datasets

Multi-Object Datasets (https://github.com/deepmind/multi_object_datasets) post-processed for easy replication of the training and evaluation of the ICML'21 paper "Efficient Iterative Amortized Inference for Learning Symmetric and Disentangled Multi-Object Representations". Datasets for CLEVR6, Multi-dSprites, and Tetrominoes are in the .h5 format and split by training and test. Images and masks are provided for each. CLEVR6 also contains the ground truth latent factors for disentanglement evaluation. Unzip the provided zip file in the desired location.本数据集为经后处理的多目标数据集（Multi-Object Datasets，https://github.com/deepmind/multi_object_datasets），可便捷复现2021年国际机器学习大会（ICML 2021）论文《面向对称与解耦多目标表征学习的高效迭代摊销推理》（"Efficient Iterative Amortized Inference for Learning Symmetric and Disentangled Multi-Object Representations"）中的训练与评估流程。其中CLEVR6、Multi-dSprites与Tetrominoes三类数据集均以.h5格式存储，并按训练集与测试集划分，每个样本均配套提供图像与掩码。CLEVR6还额外包含用于解耦表征评估的真实潜在因子。请将提供的压缩包解压至指定路径。

搜集汇总

背景与挑战

背景概述

Multi-Object Datasets是一个发布于2021年的数据集，旨在支持ICML'21论文中关于对称和解耦多对象表示学习的研究。该数据集包含CLEVR6、Multi-dSprites和Tetrominoes三个子数据集，以.h5格式提供训练和测试分割，包括图像、掩码以及CLEVR6的潜在因子，便于解耦评估。