perona-lab/cfc26

CFC26是一个大规模基准数据集，专为水下ARIS声纳视频中的鱼类检测、跟踪和计数而设计。该数据集旨在评估分布偏移下的泛化能力和部署相关性能，特别是在生态多样性和声学挑战性环境中。它涵盖多个河流系统，在鱼类大小、密度、声纳范围和背景结构方面具有显著变化，并引入了反映真实部署条件的评估设置，还支持对方向性计数准确性的分析，这对于生态监测中估计净鱼类通过量至关重要。数据集包括来自9个河流系统的11个部署位置，共4,363个剪辑，超过100万个边界框，14,072条轨道，注释格式为COCO JSON和MOT，传感器为ARIS Explorer，许可证为CC BY 4.0。分割按位置和时间顺序分配，使用录制中的自然间隙形成按时间顺序排列的非重叠训练、验证和测试集，以避免跨分割出现近重复时间上下文，并提供更现实的部署性能估计。CFC26 is a large-scale benchmark dataset for fish detection, tracking, and counting in underwater ARIS sonar video. It is designed to evaluate generalization under distribution shift and deployment-relevant performance, particularly in ecologically diverse and acoustically challenging environments. The dataset spans multiple river systems with substantial variation in fish size, density, sonar range, and background structure, and introduces evaluation settings that reflect real deployment conditions. It also supports analysis of directional counting accuracy, which is critical for estimating net fish passage in ecological monitoring. The dataset includes 11 deployment locations across 9 river systems, with 4,363 clips, over 1.0 million bounding boxes, 14,072 tracks, annotation formats in COCO JSON and MOT, sensor ARIS Explorer, and license cc-by-4.0. Splits are assigned temporally per location using natural gaps in recording to form chronologically ordered, non-overlapping train, validation, and test sets, which prevents near-duplicate temporal contexts from appearing across splits and provides a more realistic estimate of deployment performance.