Noisy-LVIS
收藏资源简介:
Noisy-LVIS是一个针对实例分割任务的大型词汇长尾数据集,包含1203个不同的类别,且含有标签噪声。该数据集由南洋理工大学创建,旨在模拟真实世界中数据的长尾分布和标签错误情况。数据集通过在LVIS v1基础上添加噪声标签生成,用于评估和改进实例分割算法在噪声环境下的性能。Noisy-LVIS的应用领域主要集中在提高模型对长尾和噪声数据的鲁棒性,解决实际应用中的挑战。
Noisy-LVIS is a large-scale vocabulary long-tailed dataset for instance segmentation tasks, which includes 1203 distinct categories and contains label noise. Developed by Nanyang Technological University, this dataset is designed to simulate the long-tailed distribution and label error scenarios present in real-world data. It is generated by adding noisy labels to the base LVIS v1 dataset, and is intended for evaluating and enhancing the performance of instance segmentation algorithms under noisy environments. The primary application scope of Noisy-LVIS focuses on improving the robustness of models against long-tailed and noisy data, as well as addressing practical challenges encountered in real-world applications.

- 1A Benchmark of Long-tailed Instance Segmentation with Noisy Labels南洋理工大学 · 2023年



