Best Computer Vision Tools

Roboflow is the quickest way to go from a folder of photos to a working vision model. It covers the whole loop: upload and label images (with Segment Anything-based tools and auto-labeling), train a model on hosted GPUs, test it, then deploy it to a hosted API or your own hardware.

Its own model family, RF-DETR, is open source. Roboflow says it was the first real-time detector to pass 60 mAP on the COCO benchmark, the research paper was accepted at ICLR 2026, and segmentation and keypoint versions followed this year. The Inference server that runs models is also Apache 2.0 and works on NVIDIA Jetson, Raspberry Pi, CPUs and GPUs, so you are not locked into Roboflow's cloud.

Pricing is published. The free Public plan gives 15 credits a month but makes your data and models public. Core costs $79/month billed yearly ($99 monthly) for three users and private data. Credits pay for labeling, training and inference, so heavy users should budget for add-ons.

Pick it if you need a custom detector in days, not months. Skip it if your images must stay private and you have no budget; train locally with Ultralytics or RF-DETR instead.

Ultralytics makes YOLO, one of the best-known families of real-time object detectors, and its newest version, YOLO26, arrived in January 2026. One Python package covers detection, instance and semantic segmentation, classification, pose, rotated boxes and depth estimation. Training on your own labeled images takes a few lines of code.

The numbers are strong for the size. Ultralytics reports 40.9 mAP on COCO for the tiny YOLO26n and 57.5 for YOLO26x, and says YOLO26n runs up to 43% faster on CPU than YOLO11n. Models export to ONNX, TensorRT, CoreML, LiteRT and OpenVINO, so they run on phones, Jetson boards and plain CPUs.

The catch is the licence. The code and the models you train are AGPL-3.0 by default. If you ship YOLO inside a commercial product without open-sourcing your own code, you need an Enterprise License, and its price is not public. The hosted Ultralytics Platform starts free with a $25 one-time credit; Pro is $29 per seat per month plus GPU time.

Pick it if you can write Python and want the shortest path to an edge-ready detector. Skip it if AGPL is a problem and you will not buy a licence; RF-DETR's core models are Apache 2.0.

SAM 3 is Meta's third Segment Anything model, released in November 2025, and it changed what a free model can do. Earlier versions outlined whatever you clicked on. SAM 3 takes a short text prompt such as "yellow school bus", or an example image, and then finds, outlines and tracks every matching object in a photo or video. This is called open-vocabulary segmentation: there is no fixed list of classes and no training step.

Meta says SAM 3 doubled the accuracy of earlier systems on its SA-Co benchmark and takes about 30 milliseconds per image on an H200 GPU, even with more than 100 objects. SAM 3.1 (March 2026) is a drop-in update that tracks up to 16 objects in one pass and, according to Meta, doubles video throughput to 32 frames per second on one H100.

It is free to download under Meta's own SAM License, which allows commercial use with some restrictions, so read it before you ship. You need a capable GPU, and it gives you masks, not a finished app.

Pick it if you need precise outlines, fast auto-labeling or video tracking. Skip it if you need a tiny model for a phone; train a YOLO or RF-DETR model instead, using SAM 3 to label its data.

OpenCV is the free, open-source library that many vision projects rely on for the plumbing: reading cameras and video files, resizing and colour conversion, filters, feature matching, camera calibration, 3D geometry and drawing results on screen. It works from C++, Python, Java and JavaScript, on hardware from servers to single-board computers.

OpenCV 5.0 shipped in June 2026 with a rewritten deep-learning (DNN) engine. Support for ONNX operators rose from about 22% in version 4 to over 80%, so many more modern models load directly, and the release adds building blocks for running language and vision-language models. It also dropped the old C API and now needs C++17, so older code may need porting. The licence is Apache 2.0.

What OpenCV does not do is train modern neural networks, label data or give you a dashboard. You bring a trained model (from Ultralytics, Roboflow or PyTorch), and OpenCV runs it and handles everything around it.

Pick it if you are an engineer building a pipeline and want no licence fees or vendor lock-in. Skip it if you want a no-code tool or a hosted API; choose Roboflow or a cloud API instead.

Multimodal AI models now handle jobs that once needed a custom-trained detector, and Google's Gemini API is the one we would try first, because its docs support object detection with bounding boxes and segmentation outlines out of the box. You send an image and a plain-English request, such as "find every dented can", and it returns labels, boxes or masks as structured data. It also reads text, answers questions about photos and understands video.

Google's docs recommend Gemini 3.8 Flash for image understanding. It costs $0.75 per million input tokens and $3.75 per million output tokens until 31 December 2026, then doubles, and there is a free tier. A small image counts as 258 tokens, so the input for one image costs a tiny fraction of a cent; the answer adds more.

The trade-offs are real. A general model is not tuned to your objects, so expect less consistent boxes than a detector trained on your own images. Each call also travels to Google's cloud, so it will not work offline or keep up with a 30-frames-per-second camera.

Pick it if you need a prototype today, handle many object types or process low volumes. Skip it if you need real-time, offline or pixel-exact results. See our Gemini 3.8 Flash page.

LandingLens is LandingAI's platform for visual inspection in factories: finding scratches, dents, missing parts and other defects on a production line. You label images, train a model and deploy it from one guided web app, so quality engineers can own the project without a machine-learning team.

Deployment fits factory floors well. Models can run in LandingAI's cloud, on site through the LandingEdge app, or in a Docker container. The free plan is usable for trials: 1,000 credits a month, up to 3 invited users, unlimited projects and 10,000 images per project. With Fast Training, training on or running one image costs 1 credit.

The limits are commercial. The free plan's single model download is for non-commercial use only, and production use needs the Enterprise plan, which has no public price. LandingAI's main pricing page now covers only its document product, Agentic Document Extraction, which suggests where the company's attention has moved.

Pick it if you run quality inspection and want a guided tool your engineers can manage. Skip it if you need general-purpose detection or published prices; Roboflow is more open about cost.

Amazon Rekognition is the simplest choice if you already run on AWS and want common vision tasks without training anything. Its ready-made APIs detect objects, scenes and landmarks, analyse and compare faces, read text, spot protective equipment and flag unsafe content in images and stored video.

Prices are low and clear. Image analysis costs $0.001 per image for the first million images a month, and new accounts get 1,000 free images a month in each API group for 12 months. Stored-video label detection is $0.10 per minute. For your own objects, Custom Labels trains a model from labeled images for $1 per training hour, then charges $4 per hour while the model is running. Leave one model on around the clock and that is about $2,920 a month.

AWS is trimming the product. On 30 April 2026, the Streaming Events and Batch Image Content Moderation features closed to new customers, though existing users can keep them.

Pick it if your images are already in S3 and you need faces, text or moderation at volume. Skip it if you need edge deployment or open-vocabulary detection; Rekognition runs only in AWS's cloud.

Google Cloud Vision is Google's long-running ready-made image API. With one call it can add labels, locate objects, read text (including dense documents), detect faces, logos and landmarks, and flag unsafe content with SafeSearch.

Pricing is simple. The first 1,000 units of each feature every month are free. After that, most features cost $1.50 per 1,000 images, object localization costs $2.25 and web detection $3.50, with discounts above 5 million a month. Each feature you ask for on an image is billed separately, so request only what you need.

The bigger picture is that Google is steering new vision work toward Gemini. At Cloud Next in April 2026, Vertex AI became the Gemini Enterprise Agent Platform. Vertex AI Vision, Google's managed video-analytics service, was deprecated on 15 June 2026 and ends on 30 September 2026, with Cloud Vision API named as one migration path. AutoML image training for custom models still exists inside Agent Platform as a separate product.

Pick it if you want dependable OCR or labels at predictable prices on Google Cloud. Skip it if you need your own object classes; the Gemini API or Roboflow is more flexible.

Clarifai built its name on image recognition and still describes a full platform for vision work: custom classification and detection, OCR, video analysis and content moderation, plus a catalogue of vision and language models you can chain together in workflows. Its Compute Orchestration product runs models on Clarifai's GPUs, your own cloud or on-premises servers.

The reason it ranks this low is uncertainty. In May 2026 Nebius hired Clarifai's core engineering team, including founder and CEO Matthew Zeiler, took its patents and licensed its inference and orchestration technology. The report of the deal did not say what happens to the Clarifai platform or its customers. When we checked on 25 September 2026, Clarifai's pricing and documentation sites did not load for us.

On pricing, Clarifai has said it retired its old self-serve plans in favour of one Pay-As-You-Go plan billed on tokens and GPU time, with a default $100 monthly spending cap. We could not confirm current rates.

Pick it if you already run production workloads on Clarifai and have a support contract that covers its future. Skip it if you are starting fresh; Roboflow or a cloud API is the safer bet today.

Azure AI Vision, now branded Azure Vision in Foundry Tools, is Microsoft's ready-made image API. Version 4.0 reads text, writes captions and dense captions, adds tags, detects objects and people, and suggests smart crops. Version 3.2 adds brands, faces, landmarks and adult-content checks. It also runs in containers, including disconnected ones, which some regulated customers need.

We rank it last because Microsoft is winding it down. The Image Analysis API (versions 3.2 and 4.0) will be retired on 25 September 2028, after which calls fail. Custom Vision, Microsoft's tool for training your own classifier or detector, retires the same day. The custom-model and background-removal features inside Image Analysis 4.0 were already switched off on 31 March 2025. Microsoft asked customers to have a migration plan in place by 25 September 2026.

Microsoft's suggested replacements are Document Intelligence for OCR, the Face API for faces, GPT models in Microsoft Foundry, and Azure Content Understanding for managed image analysis.

Pick it if you already run it in production and need a stable bridge while you migrate. Skip it for any new project; build on Foundry models, Roboflow or Ultralytics instead.