Voxel51

Vivint is transforming how millions of homeowners experience smart home security, processing massive volumes of visual data from millions of cameras deployed across doorbell, outdoor, and indoor devices to deliver real-time insights and peace of mind. Their computer vision team is pushing the boundaries of what's possible in smart home AI, building advanced models for object detection and tracking that identify people, vehicles, animals, and packages with precision. Beyond traditional security alerts, Vivint is developing next-generation features like presence-aware 'follow-me' lighting that tracks human poses and adapts to homeowners as they move through their spaces — creating truly responsive, intelligent homes. With 2M+ customers relying on these real-time insights for safety and peace of mind, Vivint's ML team operates at an impressive scale. Their data-centric approach, powered by Voxel51 FiftyOne, has been instrumental in delivering this level of accuracy and innovation: • Centralized 2TB of visual data into one unified hub and eliminated scattered tooling • Improved collaboration through shareable datasets that enable data-driven decisions • Reduced false positives and detected edge cases by gaining clear visibility into where models struggle As Chris Hall, Principal Machine Learning Engineer at Vivint shares: “FiftyOne revealed unexpected edge cases coming from window reflections, glare, and nonlinear wall proximity. It really helped us not only address this problem but also prove why a seemingly simple problem wasn’t so simple and required a more nuanced, data-driven approach.” Vivint's commitment to data-centric AI is setting a new standard for smart home technology. Read their full story: https://lnkd.in/eAAYqv3Y

🤖 🏠 🚧 Home Robots: Will Data Make or Break the Hype? 🤖 🏠 🚧 The home robot gold rush is on --- but how many of the promises will survive a showdown with messy real world data? 1X recently offered customers to pre-order the Neo home robot for $20K, and Sunday Robotics went out of stealth this week, accepting applications for its Memo home robot. Each of these collects the all-important data needed for AI training in different ways. Neo collects data obtained from early adopters using the robot in their homes, an approach reminiscent of Tesla collecting data to train its AVs. Conversely, Memo is trained on data from people performing real tasks, as opposed to synthetic data or simulations. 🚧 But significant technical challenges with loom, not to mention the immense privacy and social challenges. These challenges are also massive opportunities for innovation: - Data generalization: Can a robot really adapt to unpredictable home environments and habits?    - Data flexibility: Teaching by example (imitation learning) is a common approach in the robotics literature. But, there is little evidence non-technical users can do this effectively.    - Data protocols and ontology: How is the ocean of raw home robot data organized, annotated, and benchmarked for real long-term progress? - Data transparency: Will users have real access to and a true ability to control their home robot's data?    Don't get me wrong, these may read like criticisms, but they are seriously difficult technical challenges that are true opportunities for potential success in the space. Data is the heart of my world at Voxel51 and real human-AI teams in physically grounded settings are the focus at University of Michigan Robotics Department; so this topic is on my mind all the time. Would love to hear your thought on it, what is missing here? Do you think it's more solved than I do?

💡 🎤 🔥 Recording this Chinstrap Community fireside chat was a blast. Somehow over the course of an hour we covered many fun and interesting topics 🎤 Why and how Voxel51 is an open-source first company 🎤 What it was like spinning out a company from University of Michigan College of Engineering 🎤 What is was like funding the company via a grant vs VC funding 🎤 How to build grassroots communities around an open-source startup I'm all in on what Chinstrap is trying to achieve: create a community at the intersection of open source enthusiasts and tech entrepreneurs. Nice work team! Full video on YouTube: https://lnkd.in/eK9_ijxD Here's the Chinstrap post: https://lnkd.in/edfryZmt

Join us on Jan 14 for the Best of NeurIPS 2025 virtual event. Register for the Zoom: https://lnkd.in/d48hvUHw Talks will include: * EgoExOR: An Ego-Exo-Centric Operating Room Dataset for Surgical Activity Understanding - Ege Özsoy at Technical University of Munich * SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation - Claudia Cuttano at Politecnico di Torino * Nested Learning: The Illusion of Deep Learning Architectures - Ali Behrouz Google / Cornell University * Are VLM Explanations Faithful? A Counterfactual Testing Approach - Santosh V. at Mercedes Benz Research and Development North America *********** Want to build better computer vision models? FiftyOne is an open source toolkit from Voxel51 (our Meetup sponsor) that helps you curate datasets, evaluate model performance, visualize embeddings, catch annotation errors, and eliminate duplicate images—all in one place. “pip install fiftyone” is all it takes to get started - https://docs.voxel51.com/ #computervision #ai #artificialintelligence #machinevision #machinelearning

Voxel51 and Foretellix are partnering to solve one of the biggest challenges in AI-powered AV development: turning real-world drive logs into high-fidelity training and validation data at scale. As AV stacks move toward end-to-end architectures, teams need massive amounts of diverse, high-quality data. But real-world drives rarely capture the edge cases you need, making simulations essential. Yet poor input data creates flawed reconstructions that waste weeks of engineering time and millions in compute costs. With Voxel51's data auditing technology integrated into Foretellix's Physical AI Toolchain, teams can now seamlessly transform real-world drive logs into high-fidelity NVIDIA Omniverse 3D reconstructions and simulations. Engineering teams can scale AV development with confidence, knowing every reconstruction is built on verified, high-quality data. Read the full announcement: https://lnkd.in/eAXV9zpn

𝐀𝐧𝐧𝐨𝐮𝐧𝐜𝐞𝐦𝐞𝐧𝐭: 𝐅𝐨𝐫𝐞𝐭𝐞𝐥𝐥𝐢𝐱 𝐚𝐧𝐝 𝐕𝐨𝐱𝐞𝐥𝟓𝟏 𝐏𝐚𝐫𝐭𝐧𝐞𝐫 𝐭𝐨 𝐭𝐫𝐚𝐧𝐬𝐟𝐨𝐫𝐦 𝐀𝐈-𝐩𝐨𝐰𝐞𝐫𝐞𝐝 𝐀𝐕 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭.🚀 As the autonomous vehicle industry shifts toward end-to-end architectures, the bottleneck is not just data volume but data quality and diversity. By integrating Voxel51's data auditing technology into the Foretellix Physical AI Toolchain we enable engineering teams to turn raw drive logs into high-fidelity, variation-rich simulation datasets. This integration brings scalable 3D neural reconstruction directly into the Foretellix Foretify ecosystem. Ensure your AV development pipelines start with pristine data and end with comprehensive safety/ODD coverage. Read the full announcement: https://lnkd.in/dSJgezMW #PhysicalAI #AutonomousVehicles #ADAS #AV

Huge thank you to everyone who joined us for the #NeurIPS2025 Multimodal AI Meetup with our co-hosts at NVIDIA, Databricks, and Together AI. 🥂 It was one of those rare evenings where researchers, engineers, and builders from every corner of the multimodal ecosystem came together, and every conversation kept circling back to the same thing: 💡Good data is the foundation of good models. Hearing this echoed from so many industry leaders was energizing and a clear signal of where the field is heading. If you're working on visual AI and want to explore how data quality impacts your models, learn more: https://lnkd.in/g3Js2fu4

Solve data scarcity issue with FiftyOne and NVIDIA AI Cosmos Transfer 2.5 🦋 When your training data doesn’t match deployment conditions, your model fails in production. This is especially critical for rare or underrepresented classes where you can't afford to collect thousands of new samples. For example, moths are the rarest class in BioTrove, a biodiversity image dataset, with most images captured indoors instead of real field conditions. Models trained on this data won't recognize moths in actual crop fields where lighting, backgrounds, and context differ completely. This lab-to-field domain gap kills model performance. That’s where FiftyOne and NVIDIA Cosmos Transfer 2.5 integration comes in: 🔎Identify the issue: Use FiftyOne's semantic search to visualize class imbalance and confirm domain shift in seconds ⚡Generate realistic variations: Cosmos Transfer 2.5 preserves moth identity while transforming backgrounds into photorealistic field scenes with crops, soil, and natural lighting 🔄Curate at scale: Group original and generated images side-by-side in FiftyOne to filter out failures and build a clean dataset With this workflow, you can turn scarce, unrealistic data into balanced, diverse, training sets that match the deployment conditions, improving model performance. Try it now: https://lnkd.in/e_uZK6aB