jnzhang/hot3d-benchmark
收藏资源简介：
hot3d-benchmark数据集是从HOT3D-Clips（Aria训练集）中提取的，每个样本是一个经过修剪的子片段，覆盖单个移动对象的连续运动窗口，并包含该对象的2,000个3D和2D跟踪点以及由Molmo2-8B生成的自然语言描述。数据集包含2,534个子片段，每个子片段包括1408×1408分辨率的RGB视频、2D和3D跟踪数据以及描述信息。数据集的构建过程包括运动检测、子片段生成和重新描述等步骤。
The hot3d-benchmark dataset is derived from HOT3D-Clips (Aria train split). Each sample is one trimmed sub-clip covering a single moving objects contiguous motion window, paired with the subject objects 2,000 3D + 2D tracked surface points and a one-sentence natural-language caption produced by Molmo2-8B. The dataset contains 2,534 sub-clips, each including an RGB video (1408×1408 resolution), 2D and 3D tracking data, and caption information. The dataset construction process involves motion detection, sub-clip materialization, and re-captioning.
