Automation Overview

Auto-Label (AL) vs Custom Auto-Label (CAL)

Auto-Label and Custom Auto-Label are both very powerful tools that can be used to boost labeling efficiency. It is important to understand their similarities and differences, however, to effectively address your specific data needs:

Similarity:

• Automatically detects objects and labels them

Differences:

• Auto-Label is a pre-trained model developed by Superb AI that detects and labels 100+ common objects (cars, street signs, cans, etc), whereas Custom Auto-Label is a model trained using your own data that detects and labels niche objects.
• Auto-Label can be used “out of the box” without any pre-labeled data, whereas CAL requires at least 100 manually labeled data to be initially trained and used.

Generally speaking, if your dataset is primarily composed of common objects that Auto-Label is capable of labeling, use Auto-Label. Auto-Label will most definitely outperform CAL in this case, as Auto-Label has been trained with extensive sets of datasets by Superb AI’s veteran ML engineers.

On the other hand, use CAL if your dataset consists mostly of niche objects that AL cannot label. While Superb AI will continue to work on Auto-Label to grow the number of common objects that it can detect, it is impossible to account for each and every object - which is why we offer CAL.

If your dataset contains a mix of both common and niche objects, you can leverage both AL and CAL to maximize your labeling efficiency.

Smarter Review - Uncertainty Estimation

After AL or CAL finishes running, it informs the user how “difficult” each automated labeling task was (*red=difficult, yellow=moderate, green=easy). Higher the difficulty, more likely that AL or CAL incorrectly labeled the object.

This “difficulty”, or estimated uncertainty, is calculated based on factors such as small object size, bad lighting conditions, complex scenes, and so on. During the review stage, you can easily sort and filter labels by difficulty to prioritize going over labels with higher chance of error.

Our Auto-Label is industry-leading and CAL’s performance improves as you feed it more data, but we understand that they can’t detect and label each and every object perfectly. The Uncertainty Estimation feature takes this reality into account and helps you make the most out of your time and workforce.