Roboflow

Roboflow builds computer vision tools spanning datasets, annotation, model training, deployment and application logic. It integrates foundation models such as Meta’s Segment Anything and develops its own RF-DETR model family. Its infrastructure connects predictions to tracking, counting, camera processing and external systems—the work needed to turn visual recognition into an application. Cofounders Joseph Nelson and Brad Dwyer serve as CEO and CTO, respectively, as of September 2026.[1][7][13][9][3]

The company’s dataset and deployment tools grew out of earlier augmented reality game projects, where the founders built internal tools for recognizing physical boards and pieces.[4][1]

In his 2023 Latent Space interview, Nelson traced the collaboration to Magic Sudoku in 2017: a phone recognized a puzzle and overlaid its solution. Nelson was then running Represently, which sorted congressional messages using natural language processing; Dwyer ran the social gaming company Hatchlings. They returned to board games in summer 2019. At that year’s TechCrunch Disrupt hackathon, they added chess recognition to Board Boss in 48 hours, converting the board state into notation for the open-source Stockfish engine to recommend moves.[4]

Nelson recalled the game applications attracting attention and then losing momentum. The internal tools offered another direction: “we gotta try something different.” He dated a README manifesto for reusable vision infrastructure to November 2019 and the tools release to early 2020. Y Combinator records Roboflow as founded in 2019 and part of its Summer 2020 batch; the company dates its launch to 2020.[4][2][1]

Roboflow’s first released tool, according to Nelson, was an annotation-format converter. Different datasets and training frameworks describe bounding boxes using different conventions: corner coordinates or a center plus dimensions, absolute pixels or normalized values. A faulty conversion can silently put boxes in the wrong places while training continues. Conversion and visualization therefore addressed a small but consequential part of building a model.[4]

Another early project exposed the difficulty of annotation itself. Nelson recalled finding missing labels, including strollers, in Udacity’s academic self-driving dataset and relabeling it with Dwyer. In February 2020, they completed a rerelease on a train in Taiwan to meet a Pioneer deadline. He acknowledged that their revised annotations probably still contained mistakes.[4]

Roboflow expanded to tools for collecting, managing, labeling and processing data, training models and deploying them through hosted or self-hosted infrastructure. Dataset versions preserve images, annotations, training and evaluation splits, preprocessing and augmentation settings, providing a reproducible input for training. Developers can use the interface, APIs or a mixture of both.[2][14]

Inference runs vision models and processes camera streams. Its Workflows compose blocks with common model interfaces, allowing developers to chain models, track objects, count them, measure dwell time and add business logic. An application can process an RTSP stream, pass predictions through tracking or counting blocks, and send results to an API, webhook or notification service. Deployment options include local and edge hardware, private infrastructure and Roboflow-hosted compute.[7]

Supervision supplies reusable, model-agnostic utilities for handling detections, loading and converting datasets, counting objects in zones and drawing annotations. A box annotator, for example, accepts an image and detections and produces an image a person can inspect. Connectors support libraries including Ultralytics, Transformers, MMDetection and Inference; RF-DETR can return Supervision detections directly.[8]

Roboflow integrated Meta’s Segment Anything Model (SAM) into its labeling interface in 2023. In his April interview, Nelson demonstrated a weld mask generated with one click, compared with four clicks using Roboflow’s earlier Smart Poly tool. He also outlined how SAM could help label porch-package images for training a smaller detector for edge deployment. The proposed workflow would use foundation-model annotations as training data for a task-specific model with lower compute requirements.[4]

By October 2024, Roboflow Annotate supported SAM 2 mask generation with positive and negative prompts. In Meta’s case study, Nelson estimated that SAM and SAM 2 had been used on more than 60 million polygons, saving about 74 years of annotation time across the community. The same account reported more than 500,000 public datasets and approximately 350 million user-labeled images on Roboflow Universe.[12]

Roboflow’s November 2025 SAM 3 announcement extended the integration across Playground testing, inference APIs, local and private-cloud deployment, Workflows and annotation. SAM 3 accepts text prompts to detect and segment objects, supporting both direct inference and automatic labeling for other model architectures. Roboflow developed the surrounding tools in partnership with Meta.[13]

The company’s 2026 American Sign Language tutorial makes the limits of automatic labeling concrete. Prompting SAM 3 with “Hand” locates and segments hands, but gives every detection that generic label. Users must review the regions and assign the specific signs—such as V, R or Q—needed for downstream training. The tutorial then fine-tunes an RF-DETR segmentation model and replaces the SAM 3 block in the Workflow. Locating a hand and recognizing its sign are separate tasks.[14]

Roboflow also develops RF-DETR, a transformer model family built on a DINOv2 vision backbone. Its current package supports object detection, instance segmentation and preview keypoint detection. The research paper was authored by Isaac Robinson, Peter Robicheaux and Matvei Popov, affiliated with Roboflow, and Deva Ramanan and Neehar Peri, affiliated with Carnegie Mellon University. The project credits LW-DETR, DINOv2 and Deformable DETR as foundations.[9][10]

RF-DETR uses weight-sharing neural architecture search to derive multiple configurations from one training run. Training samples configurations with different image resolutions, patch sizes and attention windows; subsequent search also varies decoder depth and the number of object queries. Developers can select configurations along an accuracy–latency curve rather than train each model size separately. RF-DETR-Seg adds a lightweight segmentation head, pretrained using SAM 2-generated masks on Objects-365.[10]

The research team reported 60.1 AP on COCO at 17.2 milliseconds for RF-DETR’s 2XL detection variant, measuring efficiency on an NVIDIA T4 with TensorRT 10.4 and CUDA 12.4. Evaluation also covered RF100-VL’s 100 diverse datasets. To stabilize latency measurements against GPU power throttling, the team inserted 200-millisecond pauses between forward passes and advocated measuring accuracy and latency from the same model artifact. The buffered protocol measures reproducible inference latency, not sustained throughput.[10]

RF-DETR’s licensing varies by component: the open-source package and Apache-designated weights use Apache 2.0, while Plus components, including XL and 2XL detection models, use PML 1.0. Inference’s core is Apache 2.0; enterprise functionality is source-available under an enterprise license and requires an active enterprise contract for production use. Cloud-connected services require an account and API key and are metered by usage.[9][7]

In November 2024, Roboflow raised a $40 million Series B led by GV, with participation from Craft Ventures, Y Combinator, Guillermo Rauch, Jeff Dean and Amjad Masad. The founders’ announcement identified enterprise vision AI and open-source tools as investment priorities.[6][5]

Fortune reported that Pella used Roboflow to scan products for defects, Rivian used it for quality control, and BNSF Railway used it for yard inventories and safety inspections. Roboflow told the publication that more than 25,000 organizations built with its tools, including more than half of the Fortune 100, and that its open-source packages had been downloaded more than one million times in the preceding 30 days.[6]

As of September 2026, Roboflow reports more than one million developers, including engineers at more than half of the Fortune 100. Y Combinator lists a team size of 100 and a San Francisco location.[1][2]