Deep Learning Labeling Principles and Details

## I. Principle of Dataset Category Quantity Balance

Overview: In deep learning annotation, when there are multiple categories, the quantity difference between categories should not be too large. Generally, the quantity of any category should not be less than half of the category with the highest quantity. However, adjustments can be made based on specific situations. If the quantity for a particular category is genuinely too low in the entire dataset, annotate as many as possible.

Example: Assuming a sufficient dataset size, you need to annotate five categories: A, B, C, D, and E. If category A has 100 annotations, then category B must have no less than 50 annotations, and correspondingly, C, D, and E must also have no less than 50. If category E has fewer than 50, annotate as many as possible.

## II. Principle of Feature and Spatial Distinction

Overview: In data annotation, categories can be distinguished based on differences in the features of the targets to be annotated, or based on differences in their spatial positions.

Examples:

1.  Distinguishing categories based on feature differences. In the image below, there are obvious differences between categories.

![](/media/202512/2025-12-18_082759_1396970.8119583259718262.png)
2.  Distinguishing categories based on the spatial position of the target. In the images below, picking up screws and holding an electric screwdriver have the same features but are in different positions.

## III. Principle of Annotating Adjacent and Identical Targets

Overview: When targets with similar or identical features appear around the selected annotation target, they should all be annotated. During annotation, they can be labeled as the same category, or as different categories based on positional differences if needed. This prevents misidentification of adjacent unannotated targets during recognition, which could affect performance. It also avoids recognition ambiguity during training that could impact the training outcome.

Example: In the image below, with Product A - Lid, Product B - Box, and Hand positions, if only Product A - Lid is annotated, the nearby box might also be recognized as Product A - Lid, or recognition might be inconsistent. For the hand position, if only the right hand is annotated, the left hand might also be detected during recognition, or the right hand might be missed while the left hand is detected, and they might even be detected together when close.

![](/media/202512/2025-12-18_082950_6206710.7492514900176626.png)

## IV. Principle of Multi-Category Distinction

Overview: In complex scenarios, try to distinguish as many categories as possible. This helps avoid recognition confusion.

Example 1: In the image below, if 'No Product' and 'Screw Driving Position' are not annotated, a position without a product might be recognized as having a product, and the screw driving position might be recognized as a hand.

![](/media/202512/2025-12-18_083000_6544950.47726674248188516.png)

Example 2: In the image below, the project requires recognizing the target in the left image. If the object in the right image is not annotated during the entire recognition process, it might be mistakenly recognized as the 'Foam Present' detection from the left image during operation. To avoid runtime errors caused by misidentification, annotate a category like 'Box with Product' similar to the right image.

## V. Principle of Tight Fit to Target Edges

Overview: The four sides of the bounding box should fit as closely as possible to the outermost edges of the target object. Avoid leaving excessive gaps and do not cut off any part of the target.

Example: In the left image below, for 'Label Scanning', only the scanner needs to be recognized. However, the annotated purple box extends far beyond the scanner, introducing a large amount of redundant features. This causes the model to learn many irrelevant features. When scenarios similar to these irrelevant features appear, misidentifications can occur (right image below). The red box in the left image below shows the correct annotation.

## VI. Principle of Independent Annotation

Overview: Each recognition target should be annotated individually. Do not annotate multiple targets together for convenience, especially if they are touching. Also, do not omit annotating a target just because it is located within a larger target.

Example 1: In the image below, when annotating 'Product Present', each individual product should be annotated, rather than annotating three products together with one box (like the red box below). Furthermore, if project requirements dictate that all targets should be annotated, it is incorrect to omit annotating 'Product Present' just because it is located within the 'Jig Position' target.

![](/media/202512/2025-12-18_083234_0439630.27378758349145393.png)

Example 2: In the image below, when recognizing hand positions, do not annotate multiple hands together like the blue box. Instead, annotate each hand individually like the red boxes.

![](/media/202512/2025-12-18_083248_2749070.985847296383417.png)

## VII. Occlusion Handling Strategy

**Mild Occlusion (<30%):** Draw the box around the visible part, ignoring the occluded area.
**Moderate Occlusion (30%-80%):** Draw the box around the complete estimated contour (requires manually inferring the missing parts).
**Heavy Occlusion (>80%):** Do not annotate.

## VIII. Principle of Annotation Consistency

Overview: Every bounding box must correspond to a correct category label. Ensure your category list is mutually exclusive and comprehensive. The same object must be annotated with the same category across different images. Within the defined category scope, all targets that should be detected in an image must be annotated, regardless of their size, clarity, or whether they are occluded.

## IX. Principle of Prioritizing Distinct Features

Overview: From a visual perspective, observe the target to be detected. If its outline is clear, conspicuous, and its shape is sufficiently uniform throughout the dataset, select this target for annotation.

Example: In the image below, for 'Label Scanning', if you select the features inside the purple box as the annotation target, the features might not be distinct enough because the hand's shape is variable and its position changes constantly. Therefore, selecting the features inside the red box as the annotation target is more reasonable.

![](/media/202512/2025-12-18_083316_1350620.448589661158985.png)

## X. Principle of Single Annotation for the Same Target

Overview: During annotation, if the selected targets are essentially at the same location, have very little positional offset, and have highly similar features, they should not be annotated repeatedly. Also, do not annotate the same location as different categories.

Example 1: In the images below, the positions for 'Front Case Scan' and 'Component Scan' are not very different, almost the same location, and their features are quite similar. This leads to duplicate annotations of similar features at the same location. Although the features differ somewhat, there is still a certain probability of misidentification during recognition (e.g., recognizing 'Component Scan' as 'Front Case Scan'). While a larger dataset can significantly reduce the probability of misidentification, it cannot entirely eliminate it.

Example 2: The image below shows duplicate annotations for the same target, which is not allowed.

![](/media/202512/2025-12-18_083403_0113240.35768693650739725.png)

## XI. Principle of No Missed Annotations and No Selective Annotation

Overview: Inspect the image area by area, ensuring no identifiable target is missed (including targets at edges, small targets, and overlapping targets). Targets in low-light, backlit, or blurry scenes must be annotated as long as the main outline is recognizable (to enhance model robustness). Overall, there should be no incorrect annotations, missed annotations, or selective annotation.

Example 1: Compared to the left image, the right image below misses the 'Fixture' position.

Example 2: Compared to the left image, the right image below shows an incorrect annotation for the 'Fixture' position. What should be 'Fixture' was incorrectly labeled as 'Jig'.

## XII. Principle of Prioritizing Complete Target Annotation

Overview: When a selected target is occluded, the annotation should be based on its unoccluded contour, not just the visible part.

Example: In the image below, to recognize the 'Jig', the red annotation box is correct, not the yellow one.

![](/media/202512/2025-12-18_083500_4184410.19235375394574716.png)

## XIII. Blurry Target Handling Strategy

Mild Blur (Outline discernible): Annotate normally, with the bounding box fitting the visible outline.
Moderate Blur (Outline blurry but category discernible): Annotate the bounding box and label the category accurately. Do not add an extra 'Blurry' attribute.
Severe Blur (Category unrecognizable): Do not annotate.

Click here to jump to the corresponding Chinese manual page: [Deep Learning Labeling Principles and Details](/doc/308)