BenchSeg
收藏资源简介:
BenchSeg是由巴塞罗那大学等机构联合构建的大规模多视角食品视频分割基准数据集,包含来自Nutrition5k、FoodKit等公开数据集的55道菜肴场景,共计25,284帧人工标注图像。该数据集通过自由运动的半球形覆盖视频序列,捕捉360°相机轨迹下的食物形态,为评估模型在跨视角场景下的泛化能力提供严格标准。数据经过多轮双盲验证和迭代标注流程优化,主要应用于膳食评估领域,旨在解决现有方法在新视角下分割性能骤降的问题,推动视频记忆增强模型在饮食分析中的发展。
BenchSeg is a large-scale multi-view food video segmentation benchmark dataset jointly constructed by the University of Barcelona and other institutions. It encompasses 55 dish scenarios sourced from public datasets including Nutrition5k and FoodKit, with a total of 25,284 manually annotated image frames. The dataset captures food morphology across 360° camera trajectories using freely movable hemispherical coverage video sequences, serving as a rigorous benchmark for evaluating the generalization performance of models in cross-view scenarios. Optimized through multiple rounds of double-blind validation and iterative annotation procedures, it is mainly applied in the field of dietary assessment. Its core goals are to solve the problem that existing methods suffer from sharp degradation of segmentation performance under novel viewpoints, and to advance the development of video memory-augmented models in dietary analysis.
BenchSeg 数据集概述
数据集基本信息
- 数据集名称:BenchSeg: A Large-Scale Dataset and Benchmark for Multi-View Food Video Segmentation
- 主要贡献者:Ahmad AlMughrabi, Guillermo Rivo, Carlos Jiménez-Farfán, Umair Haroon, Farid Al-Areqi, Hyunjun Jung, Benjamin Busam, Ricardo Marques, Petia Radeva
- 机构:Universitat de Barcelona, Technical University of Munich
数据集内容与规模
- 数据来源:聚合了来自Nutrition5k, Vegetables & Fruits, MetaFood3D, 和 FoodKit的55个菜肴场景。
- 数据规模:包含25,284帧经过精细标注的帧。
- 数据采集方式:在自由360°相机运动下捕捉每个菜肴。
研究目的与任务
- 核心任务:多视角食物视频分割,用于饮食分析,以实现食物体积和营养的准确估计。
- 解决的问题:当前方法存在多视角数据有限以及对新视角泛化能力差的问题。
评估与基准
- 评估数据集:在现有FoodSeg103数据集上评估了20种最先进的分割模型(例如,基于SAM的、Transformer、CNN和大型多模态模型)。
- 评估场景:在BenchSeg数据集上评估了这些模型(单独及与视频记忆模块结合)。
- 主要发现:标准图像分割器在新视角下性能急剧下降,而记忆增强方法能在帧间保持时间一致性。
- 性能提升:最佳模型组合优于先前工作(例如,在BenchSeg上mAP提升约2.5%)。
相关资源链接
- 论文:https://amughrabi.github.io/benchseg
- 代码:https://amughrabi.github.io/benchseg
- 预训练模型:https://amughrabi.github.io/benchseg (8.15GB)
- 数据集:https://amughrabi.github.io/benchseg
引用信息
bibtex @article{almughrabi2026BenchSeg, title={BenchSeg: BenchSeg: A Large-Scale Dataset and Benchmark for Multi-View Food Video Segmentation}, author={Guillermo Rivo, Carlos Jiménez-Farfán, Umair Haroon, Farid Al-Areqi, Hyunjun Jung, Benjamin Busam, Ricardo Marques, Petia Radeva}, journal={arXiv preprint 2601.07581}, year={2026} }

- 1BenchSeg: A Large-Scale Dataset and Benchmark for Multi-View Food Video Segmentation巴塞罗那大学·数学与信息学系; 庞培法布拉大学·工程系; 巴塞罗那大学·神经科学研究所; 慕尼黑工业大学·摄影测量与遥感系 · 2026年



