技术
- 分析与建模 - 计算机视觉软件
- 分析与建模 - 机器学习
适用行业
- 设备与机械
- 塑料
用例
- 预测性维护
- 视觉质量检测
服务
- 数据科学服务
- 培训
关于客户
Kaleido AI 是一家总部位于奥地利维也纳的科技公司,其使命是简化复杂的技术。该公司创建的工具可以简化和加速工作流程、培养创造力并帮助其他人将新想法变为现实。它让每个人(从个人到各种规模的企业)都能接触到视觉人工智能的最新进展。 2019年,Kaleido AI推出了自动图像背景去除器remove.bg,2020年又推出了去除视频背景的在线软件Unscreen。这些工具极大地提高了用户实现目标的速度,从而导致其受欢迎程度激增。这一成功促使 Canva 在 2021 年初收购了 Kaleido AI。同年晚些时候,Kaleido AI 推出了 Designify,这是一款人工智能驱动的工具,可为个人、汽车经销商和电子商务网站等各种用户创建自动设计。
挑战
Kaleido AI 是一家总部位于维也纳的公司,致力于通过创建加速工作流程和培养创造力的工具来简化复杂的技术。该公司推出了自动图像背景去除器remove.bg和视频背景去除器Unscreen,大受欢迎并导致其于2021年被Canva收购。然而,Kaleido AI在改进其机器学习模型方面面临着重大挑战。该公司的模型需要大量高质量数据,但他们在特定的分割任务中遇到了几种边缘情况,导致模型表现不佳。收集和标记数以万计的具有多种图案、图像、背景和纹理的现实世界图像非常困难。开放数据集没有足够的此类特定类别的高质量图像。 Kaleido AI 最初依靠现实世界的数据来训练其分割模型,但这种方法非常复杂、资源密集且成本高昂。
解决方案
为了克服这一挑战,Kaleido AI 求助于 Scale AI 来生成合成数据,以提高模型在对象识别方面的性能,并提高模型预测的交并集 (IoU)。 Scale 的机器学习工程师在 Scale 的数据管理平台 Nucleus 中分析了 Kaleido AI 的样本数据和模型推理。他们发现该模型在分割具有复杂图案的图像中的对象、阴影或透明对象或场景背景中存在阴影的对象时表现不佳。 Scale 专注于这些边缘情况,并生成了 2,650 张具有不同光照、纹理和图案的合成数据图像样本。然而,这第一遍不足以显着提高模型性能。然后,Scale 团队深入研究了 Nucleus,整理数据以进一步识别这些问题边缘情况。他们还引入了将 Scale 的合成图像与 2D 空间中的真实图像进行可视化比较的功能。该分析表明,Scale 需要在合成数据分布中包含更多纹理/图案对象和更广泛的对象类型。 Scale 总共生成了 14,583 张合成图像,涵盖 12 个类别,涵盖图案、各种对象、背景和纹理。通过这些有针对性的合成数据,Kaleido AI 的 IoU 达到了 0.794。
运营影响
数量效益
Case Study missing?
Start adding your own!
Register with your work email and create a new case study profile for your business.
相关案例.
Case Study
Smart Water Filtration Systems
Before working with Ayla Networks, Ozner was already using cloud connectivity to identify and solve water-filtration system malfunctions as well as to monitor filter cartridges for replacements.But, in June 2015, Ozner executives talked with Ayla about how the company might further improve its water systems with IoT technology. They liked what they heard from Ayla, but the executives needed to be sure that Ayla’s Agile IoT Platform provided the security and reliability Ozner required.
Case Study
IoT enabled Fleet Management with MindSphere
In view of growing competition, Gämmerler had a strong need to remain competitive via process optimization, reliability and gentle handling of printed products, even at highest press speeds. In addition, a digitalization initiative also included developing a key differentiation via data-driven services offers.
Case Study
Predictive Maintenance for Industrial Chillers
For global leaders in the industrial chiller manufacturing, reliability of the entire production process is of the utmost importance. Chillers are refrigeration systems that produce ice water to provide cooling for a process or industrial application. One of those leaders sought a way to respond to asset performance issues, even before they occur. The intelligence to guarantee maximum reliability of cooling devices is embedded (pre-alarming). A pre-alarming phase means that the cooling device still works, but symptoms may appear, telling manufacturers that a failure is likely to occur in the near future. Chillers who are not internet connected at that moment, provide little insight in this pre-alarming phase.
Case Study
Premium Appliance Producer Innovates with Internet of Everything
Sub-Zero faced the largest product launch in the company’s history:It wanted to launch 60 new products as scheduled while simultaneously opening a new “greenfield” production facility, yet still adhering to stringent quality requirements and manage issues from new supply-chain partners. A the same time, it wanted to increase staff productivity time and collaboration while reducing travel and costs.
Case Study
Integration of PLC with IoT for Bosch Rexroth
The application arises from the need to monitor and anticipate the problems of one or more machines managed by a PLC. These problems, often resulting from the accumulation over time of small discrepancies, require, when they occur, ex post technical operations maintenance.
Case Study
Plastic Spoons Case study: Injection Moulding
In order to meet customer expectations by supplying a wide variety of packaging units, from 36 to 1000 spoons per package, a new production and packaging line needed to be built. DeSter wanted to achieve higher production capacity, lower cycle time and a high degree of operator friendliness with this new production line.