MINC

MINC 比以前的材料数据库大一个数量级，同时在其 23 个类别中更加多样化和良好采样。使用 MINC，我们为两个任务训练卷积神经网络 (CNN)：从补丁中分类材料，以及在完整图像中同时识别和分割材料。对于 MINC 上基于补丁的分类，我们发现性能最好的 CNN 架构可以达到 85.2% 的平均分类准确率。我们将这些经过训练的 CNN 分类器转换为一个高效的全卷积框架，并结合全连接条件随机场 (CRF) 来预测图像中每个像素的材料，实现 73.1% 的平均类准确率。我们的实验表明，拥有像 MINC 这样的大型、良好采样的数据集对于现实世界的材料识别和分割至关重要。

MINC is an order of magnitude larger than prior material databases, while being more diverse and densely sampled across its 23 categories. Using MINC, we trained convolutional neural networks (CNNs) for two tasks: material classification from image patches, and simultaneous material recognition and segmentation in full images. For patch-based classification on MINC, we found that the best-performing CNN architectures achieve an average classification accuracy of 85.2%. We converted these trained CNN classifiers into an efficient fully convolutional framework, and combined it with fully connected conditional random fields (CRFs) to predict materials for every pixel in an image, achieving an average per-class accuracy of 73.1%. Our experiments demonstrate that large, densely sampled datasets like MINC are critical for real-world material recognition and segmentation.

搜集汇总

背景与挑战

背景概述

MINC是一个大型材料识别和分割数据集，包含23个类别，比以往数据集更大且更具多样性。它被用于训练CNN进行材料分类和图像分割，分类准确率达85.2%，分割准确率达73.1%。