TL;DR — Annotation details are the small decisions about what to label and how: whether a half-visible person counts, how tight a box should be, what to do with a reflection. They follow from the model's purpose, they belong in written guidelines before labeling starts, and they decide whether a model works in production.
Do details matter in image annotation? Profoundly.
Take a model that detects people in street camera footage. A simple-sounding question: if only the lower half or only the upper half of a person is visible, should both be annotated the same way?
It depends on context. For a privacy-masking application, detecting the upper body is what enables face blurring — but when the face area is already hidden, that annotation adds nothing. For a people-counting model, every visible fragment might matter. The answer to this "simple" question changes the whole training dataset.
What is the purpose of the object detection?
Answer this before annotation begins, not after. The purpose of the detection determines:
- Which objects need to be annotated
- How precisely they need to be annotated
- Which edge cases need special rules
- What quality threshold is good enough
How do you turn the purpose into annotation guidelines?
- Write down the model's purpose and the decision it supports — this is the reference for every later rule.
- Define the classes and the rules for each — including box tightness, minimum object size and what counts as "visible".
- List known edge cases with visual examples — occlusion, truncation at the frame edge, reflections, motion blur, objects between two classes.
- Label a pilot sample and review it together — disagreements between annotators show where the guidelines are unclear.
- Iterate — after each annotation-and-training cycle, add newly discovered edge cases and use object attributes to capture the scenarios you didn't anticipate.
Why are real-world images so hard to label?
Because real images are messy. Objects are partially occluded, lighting changes, things appear at unexpected angles, and cameras capture situations nobody thought of when the guidelines were written.
Each of these "unexpected" scenarios is a potential failure mode for the trained model. The annotation team's job is to handle them consistently — which needs clear guidelines and a feedback loop between label quality and model performance. Our Safety1st case study shows what happens when the training data misses the conditions of the real site.
Is a bigger dataset enough?
No. Up to 70% of computer vision projects fail because of an overly optimistic approach to the data used in training. Teams often prioritise dataset size over quality and use augmentation to make up for missing variety — which hides the real problem.
Augmentation (rotation, flipping, colour shifts) adds variation to data you already have, but it cannot add genuinely new scenarios. A model that has never seen an occluded bicycle in training will fail on occluded bicycles in production, however much augmentation you apply. We compare the two approaches in Data Quantity vs Data Quality.
How does BoBox handle annotation details?
We write the taxonomy and edge-case guidelines with your team before production, label a sample in a paid pilot, and update the guidelines as new cases appear — with two QA passes on every batch. That way you decide what to annotate before committing to large-scale labeling that might solve the wrong problem. See the roles in our annotation team and how our pilots and pricing work.
Frequently asked questions
What are annotation guidelines?
A written specification that tells annotators exactly what to label and how: the classes, how tightly to draw boxes or masks, and rules for edge cases such as occluded, truncated, blurred or ambiguous objects — usually with visual examples.
Should partially visible objects be annotated?
It depends on what the model is for. A person-counting model may need every visible fragment labeled, while a face-blurring model only needs the upper body. Decide this from the detection purpose before labeling starts and write it into the guidelines.
What is an edge case in data annotation?
Any example the standard rules don't clearly cover — heavy occlusion, reflections, unusual angles, objects that sit between two classes. Edge cases are where models fail in production, so they need explicit rules and enough examples in the dataset.
Can data augmentation fix a dataset that lacks variety?
No. Augmentation adds variation to scenes you already have, but it cannot create scenarios the model has never seen, such as a new type of occlusion or environment. Those have to be collected and labeled.
How often should annotation guidelines be updated?
Continuously during the project. Each annotation-and-training cycle reveals new edge cases; when it does, update the guidelines and re-check earlier labels that the change affects.



