Computer vision models: find, count and check objects in images
Computer vision models find objects in an image, outline them exactly and describe a scene in words. For finding and counting, use the Ultralytics YOLO detector; for an exact outline, SAM 3; for questions about a picture, language models with vision: Claude was 1st on Arena Vision on 16 October 2026. Below are eight tools with price, license and access from Russia.
In short
- Find and count objects on a line or in a warehouse: Ultralytics YOLO
- Outline an object by its contour or by a description: SAM 3
- Describe an image and answer a question about it: Claude, 1st on Arena Vision on 16 October 2026
- A vision model from an open-weights family: Qwen, 2nd on Arena Vision on 16 October 2026
- Camera, frames and measurements with no model training: OpenCV
- Label your own photos and train a model in the cloud: Roboflow
The list
Order: Arena Vision, checked 16 October 2026.
- #1
Claude
Anthropic, USA
- What it is good for
- Reading pictures, diagrams and photos of documents, with answers in plain language.
- Why it is on the list
- Place 1 on Arena Vision as of 16 October 2026, model Claude Fable 5 (High)
- Price
- free plan; paid from $17 per month billed annually ($20 billed monthly) (16 October 2026) plans
- Russian language
- partly per the site
- Access from Russia
- not officially supported official page
- Official site
- claude.ai
- Glossary term
- Claude
- #2
Qwen
Alibaba (Qwen team), China
- What it is good for
- Questions about a picture; some models of the family have open weights and run on your own server.
- Why it is on the list
- Place 2 on Arena Vision as of 16 October 2026, model qwen3.8-max
- Price
- see the plans on the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- Official site
- chat.qwen.ai
- Glossary term
- Qwen
- #3
Gemini
Google, USA
- What it is good for
- Questions about photos and video in plain language, fast models for a stream of images.
- Why it is on the list
- Place 6 on Arena Vision as of 16 October 2026, model Gemini 3.7 Flash (High)
- Price
- free plan; paid from $4.99 per month (Google AI Plus), Google AI Pro $19.99 per month (16 October 2026) plans
- Russian language
- yes per the site
- Access from Russia
- not officially supported official page
- Official site
- gemini.google.com
- Glossary term
- Gemini
- #4
Ultralytics YOLO
Ultralytics, USA
- What it is good for
- Finding and counting objects in photos and video: defects, goods, people without a hard hat.
- Why it is on the list
- On 14 January 2026, Ultralytics released YOLO26: according to the company, recognition on a CPU became up to 43% faster. Ultralytics, Business Wire press release, 14 January 2026
- Price
- free under AGPL-3.0; cloud platform: free plan, paid from $29 per month per seat; commercial Enterprise license on request (16 October 2026) plans
- Russian language
- yes per the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- AGPL-3.0 or a paid Ultralytics Enterprise licenseYou can use it for free only if the whole project built with YOLO, including the trained weights, is published under AGPL-3.0. Closed commercial use, and even internal use, needs a paid Ultralytics Enterprise license, priced on request.
- Official site
- docs.ultralytics.com
- #5
SAM 3*
Meta, USA
- What it is good for
- An exact object outline from a click or a text prompt: area measurement, data labeling.
- Why it is on the list
- According to a Meta paper of 20 November 2025, SAM 3 doubles the accuracy of segmenting objects from a text description in images and video compared with existing systems; its training data covers 4 million distinct concepts. arXiv 2511.16719, 20 November 2025
- Price
- free; weights are granted on request on Hugging Face (16 October 2026) plans
- Official site
- ai.meta.com/sam3
- #6
Florence-2
Microsoft, USA
- What it is good for
- An open model under MIT: image captions, finding an object by a phrase, reading text in a photo.
- Why it is on the list
- The model is trained on the FLD-5B dataset: 5.4 billion annotations on 126 million images. arXiv: Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks, 10 November 2023
- Price
- free: open model, weights download without an access request (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- MIT
- Official site
- huggingface.co/microsoft/Florence-2-large
- #7
OpenCV
OpenCV.org, open-source project
- What it is good for
- Work with cameras and frames, geometry, calibration and measurements; it often runs underneath vision models.
- Why it is on the list
- On 6 June 2026, OpenCV 5.0.0 came out; as of 16 October 2026 the library’s GitHub repository has 91.2 thousand stars. GitHub, opencv/opencv, 6 June 2026
- Price
- free, open library (16 October 2026) plans
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- Apache-2.0
- Official site
- opencv.org
- #8
Roboflow
Roboflow, USA
- What it is good for
- Label your own photos, train and run a model in the cloud without your own team.
- Why it is on the list
- In November 2024, Roboflow raised $40 million in a Series B round led by GV. SiliconANGLE, 19 November 2024
- Price
- free plan: 10 credits per month; paid Core plan from $39 per month (16 October 2026) plans
- Russian language
- no per the site
- Official site
- roboflow.com
In short: the task and the model
Computer visioncomputer vision. Computer vision is the field of AI in which programs understand images and video: they find objects, read text, count and measure. solves three different tasks, and each has its own model. Find and count: a detector draws a box around each object of the needed class. Outline and measure: a segmentation model gives an exact contour, from which area and size are calculated. Understand the scene: a language modellarge language model. A large language model (LLM) is a neural network trained on a huge amount of text. It writes, summarizes and analyzes text on request. with vision answers a question about the image in words.
Only the language models have a rank. Arena runs open boards based on blind pairwise votes from users, and Arena Vision compares exactly the models that answer questions about a picture. YOLO, Florence-2, OpenCV and Roboflow are not on that board, so their cards show a fact from a paper or a release with a date.
Finding objects
A detector finds all objects of the given classes in a frame and returns a box, a class and a confidence score. This is how you count boxes on a pallet, find people without a hard hat or spot a scratch on a part. Ultralytics YOLO runs on a camera, an industrial computer or a server. Version YOLO26 came out on 14 January 2026: according to the company, it runs up to 43% faster on a CPU.
The key point about YOLO for a business is its AGPL-3.0 license. Free use is allowed only if you open the whole product with YOLO, including the trained weights, under the same license. A closed product, and even an internal plant system, needs a paid Ultralytics Enterprise license, priced on request.
In plain words: you cannot train YOLO on your photos of defects and put it on the line without opening the code and weights, unless you buy a license. If you do not want to open them, budget for the license or take a model with a permissive license (MIT, Apache-2.0) and compare accuracy on your own images.
The launch order is the same for any detector: collect images from your line, labeltarget variable. A target variable is the value a machine learning model learns to predict: whether a customer will leave, how much product will sell, how much a flat costs. them, fine-tune the model, test it on frames it has not seen, and only then put it on the camera. Roboflow offers labeling and training without your own team: the free plan includes 10 credits a month, and the paid Core plan costs from USD 39 a month.
Outlining and measuring
A box says where an object is but does not show its shape. Measurements need a mask, an exact contour. SAM 3 (Segment Anything Model 3) came out in November 2025 with open weights: the model outlines objects in photos and video, including by a text description.
According to the developers’ paper of 20 November 2025, SAM 3 is twice as accurate at outlining objects from a text description as existing systems, and its training settraining dataset. A training dataset is the part of your data that a model uses to set its weights: examples with known answers that it learns from. holds 4 million unique concepts. In practice, you can write “all cracks” or “all green apples” and get masks without manual labeling.
SAM 3 has its own license, the edition of 19 November 2025, and the weights are given on request on Hugging Face. Before a commercial launch, a lawyer reads its terms as carefully as the YOLO terms.
Florence-2 from Microsoft handles nearby tasks with one model: image captions, finding an object by a phrase, reading text. It was trained on the FLD-5B set: 5.4 billion annotations on 126 million images. The MIT license allows commercial use without opening your own code.
Models that understand a picture
A language model with vision answers a question about an image in plain language: what is in the photo, whether there is damage, what is written on the nameplate. On Arena Vision, a user asks a question with a picture, sees two answers without model names and votes for one of them. The term is explained in the glossary: vision-language modelvision language model. A vision language model (VLM) is a neural network that takes an image and text and answers in words: it describes a photo, reads a document, or answers questions about a picture..
Places of the candidates on Arena Vision: Claude, 1st on Arena Vision on 16 October 2026; Qwen, 2nd on Arena Vision on 16 October 2026; Gemini, 6th on Arena Vision on 16 October 2026. The model versions that took these places are listed in the cards.
Access matters more than rank. Claude and Gemini do not officially support Russia. Some Qwen models are open: in August 2026, Alibaba opened the weights of a Qwen-Max class model for the first time, Qwen3.8 with 2.4 trillion parameters. You run an open modelopen weights model. An open weights model is a neural network whose parameter files you can download and run on your own hardware. on your own server, and images never leave the company.
Where the limit lies: a language model describes a single image in detail, but for a continuous camera stream it is slow and expensive. The detector handles the stream, and the language model is called for rare disputed frames when an answer in words is needed.
Faces and biometrics: the law
A face by which a system identifies a person becomes biometric personal data. Article 11 of 152-FZ, Russia’s personal data law: information that describes a person’s physiological and biological features is processed only with written consent. The exceptions are listed in Part 2: justice, defense, security and other cases where the state acts.
For a store, an office or a warehouse, the conclusion is practical. Identifying employees at the entrance requires written consent from each of them. Counting visitors and measuring a queue can be done with a people detector that does not compare faces with a database; a lawyer still checks the setup before launch.
Storage is regulated too: personal data of Russian citizens is collected and stored in databases located in Russia. So frames with faces are not sent to a foreign cloud service, and the model runs on the company’s own server. The term is explained in the glossary: face recognitionfacial recognition. Facial recognition is a computer vision technology that finds a face in a photo or video and compares it with samples to identify a person or confirm who they are..
Vision in manufacturing
In a factory, vision pays off where a person looks today: defect control, counting, label checks, a hard hat and a vest in a danger zone. How a steel plant checks its products with computer vision and predictive modelspredictive analytics. By definition, predictive analytics is the calculation of the probability of future events from historical data: demand, equipment failure, customer churn or late payment. is covered in the article about quality control at Severstal.
Pilot steps: choose one area with a clear cost of error, collect images of defects and normal parts, train a detector and compare its decisions with the inspector’s for a week. Ready scenarios for plants are on the page AI for manufacturing.
The vision server sits on site: frames do not leave the perimeter, and the model works without the internet. OpenCV under Apache-2.0 and Florence-2 under MIT fit this with no terms about opening code, while YOLO needs a license decision before you buy cameras.
When vision needs hands, robotics begins: a manipulator picks up the part the camera found. Models and robots for this are collected in the overview of robotics.
How we built this top
- The language model’s place on Arena Vision on the check date
- For detectors and libraries: a fact from the developer’s paper or release, with a date
- License: whether closed commercial use is allowed
- Running on your own server and access from Russia
Order: Arena Vision. Places go only to tools whose models are on the board.
Checked:
How we build ratingsQuestions and answers
Which neural network recognizes objects better?
Can you recognize visitors’ faces?
Sources
- Arena, March 2026 update, March 2026arena.ai
- InfoQ, 26 November 2025infoq.com
- Pravo.ru, ruling of Moscow's Tverskoy District Court of 21 March 2022 (does not apply to WhatsApp or to users), 21 March 2022pravo.ru
- ConsultantPlus, 152-FZ, Article 11, 26 July 2026consultant.ru
- Arena Vision, checked 16 October 2026arena.ai
* Meta is recognized as an extremist organization and its activity is banned in Russia.