Understand the training platform
The workstation uses a detection model to recognize parts. The training platform is where you make that model. This page covers the whole path from footage to model, and a few terms you will see often on the platform.
These pages are for process engineers who build SOP processes and train models. For how to build a process and submit it for training in the admin console, see Train and choose a model. For how to record footage and verify a model on the workstation, see Understand sampling for training.
From footage to model
This is how the model for an operation is made:
- Submit for training: In the admin console, click 送训 (Submit for training). The training platform creates the project automatically, and the operation's teaching video is synced over automatically.
- Footage: Videos and images. Besides the teaching video, you can record footage on the workstation or upload it directly on the training platform.
- Extract frames: Cut the video into individual images, one frame at a time.
- Auto-labeling: AI first draws boxes around the parts in each image.
- Labeling and review: A person checks the boxes the AI drew and confirms them or fixes them. This step takes the most time, and it affects accuracy the most.
- Training: Train a model with the reviewed images. When training finishes, the model is exported automatically to a format the workstation can run.
- Release a test version: First release it to the workstation and verify it for real.
- Release a stable release: When verification is good, release the stable release. All workstations switch to the new model automatically.
- Hard samples: Images the workstation got wrong (false alarms, missed detections) are sent back to the training platform. After you confirm them, they go into the next training run. Then you go back to step 6.
Some terms
| Term | Meaning |
|---|---|
| Project | One operation has one project. It is created automatically when you submit for training from the admin console. Its name is "product-operation-date-number" |
| Class / label name | The thing the model must recognize, for example pen_tip. The label name is in lowercase English. The Chinese name is set in the admin console |
| Frame | One image cut from a video |
| Labeling / box | Drawing a box around a part in the image and marking which class it is |
| Review | A person has checked the boxes the AI drew and confirmed them or fixed them. Only reviewed frames are the most reliable |
| mAP50 | Model accuracy, from 0 to 1. Higher is better. Above 0.9 usually means the model recognizes things very accurately |
| Test version / stable release | The test version is released to the workstation first for a trial. When the trial is good, release the stable release, and all devices switch to it automatically |
| Hard samples | Images the workstation got wrong (false alarms, missed detections), sent back to add to training |
This page describes the live version of 2026-10-22
Names such as the "帧 / 已核查" (Frames / reviewed) column in the project list and the "标签名 / 检测短语" (label name / detection phrase) columns in the class table only exist in this version. If your screen uses different names, the platform has not been updated to this version yet.