COCO-O
收藏资源简介:
COCO-O是一个基于COCO的大型测试数据集,专门设计用于评估对象检测器在自然分布偏移下的鲁棒性。该数据集包含6782张在线收集的图像,涵盖6个测试域:素描、天气、卡通、绘画、纹身和手工艺。COCO-O旨在提供一个更全面的鲁棒性评估,与训练数据有较大的分布差异,导致Faster R-CNN检测器性能相对下降55.7%。数据集的应用领域包括自动驾驶等实际场景,旨在解决模型在面对环境变化时的性能下降问题。
COCO-O is a large-scale test dataset based on COCO, specifically designed to evaluate the robustness of object detectors under natural distribution shifts. This dataset contains 6,782 images collected online, covering 6 test domains: sketches, weather conditions, cartoons, paintings, tattoos, and handicrafts. COCO-O aims to provide a more comprehensive robustness evaluation, as its data has a significant distribution gap compared to training data, leading to a relative performance drop of 55.7% for Faster R-CNN detectors. The application scenarios of this dataset include real-world use cases such as autonomous driving, and it is intended to address the issue of model performance degradation when facing environmental changes.

- 1COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts阿里巴巴集团 · 2023年



