How we guarantee height
accuracy level
Six steps for perfect bounding box annotation.
Depending on the complexity, from 3 to 12 people work on every single image. Thus, we are able to achieve the annotation quality our customers need.
We have developed the labeling software tool to fit our workflow, where there are three roles involved in the annotation process:
- annotator
- verifier
- inspector/project manager
STEP 1. The Manager receives the customer's task, clarifies details, and prepares a technical job for the annotation team.
The task:
"We need to label shampoo bottles with bounding boxes and classify them by brand name. Bounding boxes have to be drawn along the border without "cuttings" and extra spaces. A given object shouldn't be annotated if it's less than 50% visible in the picture."
The technical job prepared by a manager:
It is required to label each bottle of shampoo STRICTLY along the border using bounding boxes. The verifier examines each image and, if there is any extra space or "cropping, sends it back for correction. Each item must be classified into categories according to brand names.
Possible Mistakes:
The task:
"We need to label shampoo bottles with bounding boxes and classify them by brand name. Bounding boxes have to be drawn along the border without "cuttings" and extra spaces. A given object shouldn't be annotated if it's less than 50% visible in the picture."
The technical job prepared by a manager:
It is required to label each bottle of shampoo STRICTLY along the border using bounding boxes. The verifier examines each image and, if there is any extra space or "cropping, sends it back for correction. Each item must be classified into categories according to brand names.
Possible Mistakes:
- Cropped Object
- Extra Space
- Incorrectly Classified
- Undesired Label
- Unlabeled Object".
STEP 2. The annotator receives an image for labeling. Once an image is done, the annotator passes it to the verifier.
STEP 3. Verifier checks and sends the image back to another annotator if any mistakes occur.
In this case, the "Extra Space" and "Unlabeled Object" errors took place (two bounding boxes with red filling on the
picture below).
In this case, the "Extra Space" and "Unlabeled Object" errors took place (two bounding boxes with red filling on the
picture below).
STEP 4. Then the image with approved labels and comments goes to any available annotator.
STEP 5. Verifier approves the picture or repeats the
third step.
The second, third and fourth steps are repeated until the verifier finally approves all the labels on the image.
STEP 6. Finally, the manager checks out all verified images.
If the manager rejects the picture, it goes back to the verifier to step three. The process continues until the manager approves all the images from the dataset. Once the last photo from the package is approved, a JSON file is sent to the customer.
STEP 5. Verifier approves the picture or repeats the
third step.
The second, third and fourth steps are repeated until the verifier finally approves all the labels on the image.
STEP 6. Finally, the manager checks out all verified images.
If the manager rejects the picture, it goes back to the verifier to step three. The process continues until the manager approves all the images from the dataset. Once the last photo from the package is approved, a JSON file is sent to the customer.
The annotation process is not very difficult, but the process is very resource consuming and requires multiple checks. The workflow must be accompanied by software development.
We are improving our process and can customize our labeling tool if our clients ask us for non-standard tasks. If you need cost-effective and high-quality image annotation, please fill the form on our contact page or email us at hello@tagias.com.
We are improving our process and can customize our labeling tool if our clients ask us for non-standard tasks. If you need cost-effective and high-quality image annotation, please fill the form on our contact page or email us at hello@tagias.com.
Case Studies
CASE
