aviadcohz/Detecture_Benchmarking

Detecture Benchmarking Suite 是一个用于VLM引导的多纹理分割的五数据集基准套件，与Detecture架构（通过多重接地的VLM引导多纹理分割）一起发布。该套件包含训练集、一个域内测试集和三个域外评估基准，用于在Detecture论文中对Detecture与基线模型家族（SAM 3 vanilla、Grounded-SAM 3、Sa2VA和Qwen2SAM零样本）进行评分。数据集包括ADE20K_textured_images（训练集）、ADE20k_Detecture（域内测试集）、CAID（域外水/航空图像测试集）、RWTD（域外真实世界纹理对测试集）和STLD（域外合成Brodatz测试集）。每个数据集包含图像、掩码、纹理掩码、覆盖图和元数据文件。数据集采用CC-BY-4.0许可，使用时需引用相关论文和原始数据源。

The Detecture Benchmarking Suite is a five-dataset benchmark suite for VLM-guided multi-texture segmentation, released alongside the Detecture architecture (VLM-guided multi-texture segmentation via multiplexed grounding). This bundle contains the training set, one in-domain test set, and three out-of-domain evaluation benchmarks used to score Detecture against baseline model families (SAM 3 vanilla, Grounded-SAM 3, Sa2VA, and Qwen2SAM zero-shot) in the Detecture paper. The datasets include ADE20K_textured_images (training set), ADE20k_Detecture (in-domain test set), CAID (out-of-domain water/aerial imagery test set), RWTD (out-of-domain real-world texture pairs test set), and STLD (out-of-domain synthetic Brodatz test set). Each dataset contains images, masks, texture masks, overlays, and metadata files. The dataset is released under CC-BY-4.0 license and requires citation of the relevant paper and original data sources when used.