View projects and the overview
The project list shows the projects for all operations and their review progress. Open a project's "Overview" to see the class setup, adjust frame extraction settings and set up automatic retraining.
Project list
- Four numbers at the top: total projects, bound to SOP, VLM labeling, total classes.
- Search box: you can search by project name, SOP ID or SOP code.
- Columns: project, labeling backend, platform SOP ID, SOP code, frames / reviewed, number of classes, created time, actions (详情 (Details) / 编辑 (Edit) / 删除 (Remove)).
- SOP code: the code you can find under SOP 管理 (SOP Drafts) in the admin console, for example SOP-20261026-001. Several operations under the same SOP show the same code. If the platform never pushed it, or the operation has been deleted, it shows "—".
- Frames / reviewed: for example "2172 / 2163" = 2172 frames extracted, and 2163 frames reviewed by a person. If reviewed is 0, it is shown in orange. This means all the boxes in this project were drawn by AI and nobody has checked them yet. Hover over it to see the percentage. Click the column header to sort by the number reviewed. For a project that has just finished frame extraction, the numbers update after 1 minute or after you refresh.
- Top-right corner: 刷新 (Refresh), 回收站 (Recycle bin), + 新建项目 (+ New project).
You usually do not need to create projects by hand
Projects submitted for training from the admin console are created automatically. Create a project by hand only for experiments, or for global classes that are not tied to an operation.
A deleted project goes to the recycle bin first. In the recycle bin you can 还原 (Restore) it or 彻底删除 (Delete permanently) it.
Overview
Click a project's 详情 (Details). It opens on the 概览 (Overview) tab by default.
Four numbers at the top: number of classes, frames / reviewed, total boxes (including the manual edit rate), hard samples to handle (click it to go to the Hard samples page).
Class setup table
The columns are ID / label name / detection phrase (prompt) / description / threshold / box count.
- Label name: the English name used for model training and detection rules.
- Detection phrase: an English phrase that tells the AI what to look for during auto-labeling.
- Description: extra notes for the AI (what it looks like, where it is).
- The more specific the detection phrase and description, the more accurate the auto-labeling.
For projects submitted from the platform, you cannot change classes here
These projects are marked "只读 · 平台下发" (Read-only · from platform). To change classes, go to "③ 模型与类别" (③ Model & classes) of the operation in the admin console. After the change, submit for training again.
For projects created by hand, you can change them directly: click + 添加类别 (+ Add class), and when done click 保存类别 (Save classes).
- If you save with an empty label name, you see "第 N 行标签名不能为空" (The label name in row N cannot be empty).
- "混合类别" (Mixed classes) is a reserved name and cannot be used as a label name.
Sometimes "疑似改名的旧素材" (Old footage that may have been renamed) appears below. This means a class was renamed on the platform, but old footage is still attached to the old name. Click claim to move it over.
Project parameters
They are on the right. Click 编辑 (Edit) to change them.
| Parameter | Meaning | Recommendation |
|---|---|---|
| Image quality jpeg_q | Quality of the extracted images. 2 is the sharpest | Keep 2 |
| Frame width frame_width | Width of the extracted images | 1920 |
| Part fps / SOP fps | How many images to extract per second | Usually 2. For operations with fast movements, you can raise it |
| Validation ratio val_ratio | How many images to keep for testing the model | Default 0.15. Usually no need to change |
| Detection threshold box_threshold | How confident auto-labeling must be before it draws a box | Default 0.15 |
After you change the frame extraction fps or frame width, extract frames again for it to take effect
The validation ratio and detection threshold take effect at the next training or labeling run.
Closed-loop settings
- Automatic retraining threshold: when this many newly confirmed hard samples have built up, and no other training is running at the time, training runs again automatically. Default 0 = off.
- VLM container class / container prompt: when parts sit in a bin or tray, this tells the AI to find the container first and then look for parts inside it. Usually you do not need to set it.