VISAT
收藏资源简介:
VISAT是一个新型的开放数据集,旨在评估交通标志识别中视觉模型的鲁棒性。该数据集基于Mapillary交通标志数据集(MTSD)构建,引入了两个基准测试,分别强调对对抗性攻击和分布偏移的鲁棒性。数据集包括了80%的训练数据、10%的验证数据和10%的测试数据,共计303929条数据。VISAT数据集为每个交通标志创建了额外的视觉属性标签(颜色、形状、符号和文本),以扩展其在评估多任务学习模型鲁棒性和识别多任务学习任务之间的虚假相关性方面的能力。
VISAT is a novel open dataset designed to evaluate the robustness of visual models in traffic sign recognition. Constructed based on the Mapillary Traffic Sign Dataset (MTSD), this dataset introduces two benchmark tasks that respectively emphasize robustness against adversarial attacks and distribution shifts. It includes 80% training data, 10% validation data, and 10% test data, with a total of 303,929 samples. The VISAT dataset provides additional visual attribute labels (color, shape, symbol, and text) for each traffic sign, expanding its capability to evaluate the robustness of multi-task learning models and identify spurious correlations across multi-task learning tasks.

- 1通过伊利诺伊大学香槟分校, 布朗大学, 威廉与玛丽学院 · 2025年



