---
title: "Box Suggestions"
description: "Box suggestions in LumiScan Annotation Web: dashed suggested boxes from your trained network or a detector, accepted by correcting them or with Enter."
product: "LumiScan Annotation Web"
component: "lumiscan-annotation-web"
version: "2.2.0"
display_version: "2.2.0.0"
latest_version: false
language: "en"
url: "https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/box-suggestions.html"
---

# Box Suggestions – LumiScan Annotation Web

**Suggest boxes** (`P`) asks the box suggestion service for boxes on the open image. The button is only shown when the server has a service (see [Box Suggestion Service](https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/box-suggestion-service.html)), or runs the networks itself, as the [single-PC installation](https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/single-pc-installation.html#box-suggestions-without-a-server) does.

The suggestions come from:

- **your trained LumiScan network** for the project (chosen in the [project settings](https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/project-settings.html#box-suggestions)), or
- for parts without a network yet, an **open-vocabulary detector** that looks for the objects named by the labels, for example "bracket" or "hexagon nut".

![Suggested boxes, dashed, with label and confidence, for example bracket 97 %, nut 93 %, washer 74 %](https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/_images/screens/suggestions.jpg)

## Working with suggestions

Suggestions are drawn **dashed** with their label and confidence. They are not part of the annotation until you accept them.

[Video](https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/_images/videos/box-suggestions.mp4)

|   |   |
| --- | --- |
| Accept one | Correct it (move, resize, rotate, change the label), or select it and press `Enter` |
| Delete one | Right-click it, or select it and press `Delete` |
| Accept all remaining | **Accept all** |
| Delete all remaining | **Discard** |
| Save with open suggestions | You are asked whether to accept them or leave them out |

Suggestions are never written to the annotation file by themselves.

What is not suggested:

- objects that already have a box on the image (overlap with an existing box),
- objects outside the region of interest,
- suggestions below the project’s minimum confidence, and beyond its maximum number of boxes.

## Good suggestions

- **With a network**: the more annotated images it was trained on, the better. Retrain after every few hundred corrected images.
- **With the detector**: label names say what to look for. Use visible properties ("hexagon nut", "weld seam", "black rubber seal"), not internal part numbers. Labels with a meaningless name get a description in the project settings, for example `part_a: hexagon nut`.
- Tiny or low-contrast objects: lower the minimum confidence.
- Always check every accepted box. A wrong box in the training data teaches the network the mistake.
- Let the network say where it is unsure: [Rate with network](https://docs.hdvisionsystems.com/en/lumiscan-annotation-web/2.2.0/finding-and-sorting-images.html#rate-with-network) sorts the images so the most useful ones are annotated first.