当前位置:网站首页>CVPR 2022丨特斯联AI提出:基于图采样深度度量学习的可泛化行人重识别
CVPR 2022丨特斯联AI提出:基于图采样深度度量学习的可泛化行人重识别
2022-06-30 15:34:00 【智源社区】
近日,特斯联科技集团首席科学家邵岭博士及团队提出了一种高效的小批量采样(mini-batch sampling)方法——图采样(Graph Sampling, GS),用于大规模深度度量学习,极大改善了可泛化行人重识别。目前,该研究成果(题为: Graph Sampling Based Deep Metric Learning for Generalizable Person Re-Identification)已被今年的 CVPR 接受并发表。
行人重识别是一项热门的计算机视觉任务,其目标是通过对大量图库图像进行检索,以便找出给定的查询图像中的行人。在过去的两年中,可泛化行人重识别因其研究和实用价值而受到越来越多的关注。这类研究探索学习行人重识别模型对于未见过的场景的可泛化性,并采用了直接的跨数据集评估来进行性能基准测试。
目前较热门的深度学习行人重识别模型的方法包括分类(使用ID loss)、度量学习(使用pairwise loss或 triplet loss),以及它们的组合(例如ID + triplet loss)。ID损失函数对于分类学习来说十分便捷。然而,在大规模的深度学习中,涉及分类器参数会在前向和反向传播过程中产生大量的内存和计算成本。相似地,在全局视图中涉及用于度量学习的类别相关参数也效率不高。

图1:两种不同的采样方法:(左侧)PK采样器;(右侧)邵岭博士团队提出的GS采样器。不同的形状表示不同的类别,而不同的颜色则表示不同的批次(batches)。GS为所有的类别构建一个图,并且总是对最近的相邻类别进行采样

边栏推荐
- A. Theatre Square(codefore)
- 【时序数据库InfluxDB】Windows环境下配置InfluxDB+数据可视化,以及使用 C#进行简单操作的代码实例
- 4.4 string
- 分布式机器学习:模型平均MA与弹性平均EASGD(PySpark)
- Operator%
- Single cycle CPU of the design group of West University of Technology
- Advanced C language - pointer 3 - knowledge points sorting
- Experiment of the planning group of the West University of technology -- pipeline CPU and data processing Adventure
- Using asp Net core creating web API series
- Complement (Niuke)
猜你喜欢

Imeta | Ye Mao / Shi Yu reviewed the dynamic shuttle and ecological function of intracellular and extracellular genes in the environmental microbiome

The sound network has fully opened the real-time transmission network sd-rtn, which has been free of network wide accidents for seven years - this is FPA!

Three types of technical debt that programmers often encounter: code, data, and architecture

4.3 variables and assignments

The short video and live broadcast incubation training camp with goods opens nationwide enrollment!

Using asp Net core creating web API series

Rte2021 review of the practice and the way of AI OPS landing

Deep understanding Net (2) kernel mode 1 Kernel mode construct event event

Tetris source code (color version)

iMeta | 叶茂/时玉等综述环境微生物组中胞内与胞外基因的动态穿梭与生态功能...
随机推荐
Mysql事务/锁/日志总结
消息队列十连问
Xiao Sha's pain (thinking problem)
Practical cases of data visualization (timeline rotation diagram, streamlit control year metabase visualization tutorial) 2.0
Compare whether the two arrays are the same
001 data type [basic]
BYD is more and more like Huawei?
【子矩阵数量统计】CF1181C Flag子矩阵数量统计
Don't fight for big companies
Container common commands
ADB devices cannot detect the problem of Xiaomi note 3
The crystal ball "data insight" was officially launched: insight into the change of consumption trend and the details of interactive experience
Expressions in if statements
B. Moamen and k-subarrays (codeforce+ binary search)
Rte2021 review HDR technology product practice and exploration
Bye civil engineering, hello CS, can you change the certificate to the Blue Bridge Cup
Infrastructure is code. What are you talking about?
A little idea about big experiment data
How should we understand the variability of architecture design?
openresty 内置变量