DenseMarks 数据集
收藏资源简介:
DenseMarks 数据集是一个用于学习人类头部图像的规范嵌入表示的数据集。该数据集包含大量的野外头部视频,并通过现成的点跟踪器进行了标注。数据集旨在训练一个网络,该网络能够为每个像素预测一个3D嵌入,从而将2D头部图像映射到一个语义感知的3D规范单位立方体中。数据集的创建过程涉及了对比损失和多任务学习,以实现匹配点之间的紧密嵌入。DenseMarks 表示可用于寻找共同语义部分、头部跟踪和立体重建等领域。
The DenseMarks dataset is a collection dedicated to learning canonical embedding representations of human head images. It contains a large volume of in-the-wild head videos, which are annotated using off-the-shelf point trackers. The dataset is designed to train a network that predicts a 3D embedding for each pixel, thereby mapping 2D head images into a semantic-aware 3D canonical unit cube. The construction of this dataset employs contrastive loss and multi-task learning to achieve tight embeddings between corresponding points. The DenseMarks representations can be applied to tasks such as discovering shared semantic parts, head tracking, and 3D reconstruction.




