Cameras are usually already installed on site, but nobody watches the footage: human attention lasts half an hour into a shift. Vision models cover exactly the monotonous part, and they do it in several steps. First find objects in the frame, then outline the defect area properly, then describe what is visible, and only then turn it into a report. Splitting the work lets you set thresholds separately at each stage instead of waking the supervisor over every shadow.
A detection model draws boxes around what you care about: a part, a box, a pallet, a person in a restricted zone, a helmet. It runs fast and on site, without sending video anywhere. Some of these models search by text description, which helps when there are many classes and they keep changing. Without this part there are no counters and no triggers, and the rest of the logic has nothing to stand on.
The most widely used real-time object detector: finds and marks items in video even on modest hardware. YOLOv5 came out back in 2020; the catalog starts from YOLOv8.
Real-time object detector, an open alternative to YOLO without AGPL. Supports segmentation (object outlines) and, since 2026, keypoints.
Finds any objects in an image from a text description, without training on your data: "red box", "person without a hard hat". Rex-Omni is the new VLM-based generation.
Segmentation traces an object by its contour rather than a rectangle, which matters for defects: a chip, a run or a gap needs measuring, not just noticing. Depth estimation adds a sense of distance when there is only one camera. Without this part the system can only say a part is in frame, not what is wrong with it, so it never reacts to a defect.
Selects any object in photos and videos with a click or a box. The basis for background removal and object counting.
Estimates depth, the distance to every point, from one ordinary photo or video. DA3 reconstructs scene geometry from several frames.
A model that understands images and text turns a trigger into a human sentence: what the object is, what is wrong, how confident the call is. This part connects vision to reporting, because a supervisor needs words rather than box coordinates. Without it the log holds class numbers and frames that only the person who set the system up can interpret.
One of the strongest open vision models: reads documents, tables, charts and video, and works with user interfaces. Since Qwen3.5, vision is built directly into the main Qwen model.
A very small vision model: captions, object detection, segmentation and text reading from a single prompt. Runs even on a CPU.
A small, fast vision model for product use cases: answering questions, finding and pointing to objects, captions. Moondream 3.1 is a 9B MoE with 2B active.
A language model condenses a shift worth of triggers into short text: how many, where, what repeats, what to look at first. Instead of a thousand notifications the manager gets one summary, while genuinely urgent events still go out immediately. Without this part the system floods a chat, people stop reading it within a week, and the rollout quietly dies.
A family of language models with strong Russian language support, from small versions for a laptop to a flagship on par with commercial APIs.
Compact Google models that run well on a single computer; larger versions understand images. Includes CodeGemma for code, FunctionGemma 270M for function calling and the fast DiffusionGemma.
Cameras and lighting decide the outcome, not the choice of model. Glare, backlight, a dirty lens or steam wipe out any accuracy claimed on paper. Before rolling out, record your own conditions including the night shift, and gather a set of defect frames, which are always scarce. Agree up front what happens on a trigger: stop the line, tag the item, or call a supervisor. Video monitoring of employees is regulated, described here in general terms, with a lawyer reviewing your case.
An open model can run on your own server: data stays in-house, there is no per-request fee, and the model can be fine-tuned on your documents.
