CFC-DAOD

该数据集名为CFC-DAOD，是加州理工学院鱼类计数数据集（CFC）的扩展版本，旨在研究在渔业监测背景下，特别是针对在声纳视频中检测和计数迁移中的鲑鱼，开展领域自适应目标检测（DAOD）。该数据集包含了一个监督式的肯纳训练集（70,000张图片中的132,000个注释），以及一个用于领域适应方法的非监督训练集。与现有的DAOD基准相比，其规模大幅增加，包含了29,000帧、150个视频中的168,000个边界框注释。该数据集的任务是领域自适应目标检测。

The dataset is named CFC-DAOD, an extended version of the Caltech Fish Counting (CFC) dataset. It is designed to explore Domain Adaptive Object Detection (DAOD) for fisheries monitoring applications, specifically for detecting and counting migrating salmon in sonar videos. This dataset comprises a supervised Kenna training set with 132,000 annotations across 70,000 images, as well as an unsupervised training set for domain adaptation methods. Compared with existing DAOD benchmark datasets, its scale has been greatly expanded, encompassing 168,000 bounding box annotations from 29,000 frames across 150 videos. The core task of this dataset is domain adaptive object detection.

搜集汇总

背景与挑战

背景概述

CFC-DAOD是一个新的领域自适应目标检测（DAOD）基准数据集，旨在评估模型在多样化的真实世界数据上的性能。它作为ALDI项目的一部分被引入，用于解决现有DAOD基准中的系统性问题，如性能高估和缺乏通用性，并支持ALDI++方法在多个基准上实现最先进的结果。