Using EyePop.ai’s AutoLabeler

As your data uploads, EyePop.ai’s AutoLabeller begins identifying and prioritizing images which are most dissimilar from the dataset for human review. This simplifies and accelerates labeling by limiting the number of images that require your review.

Once data is uploaded, EyePop.ai directs you into prompt testing. You’ll see a set of four images on the right hand side of the screen. You can hit the refresh button underneath the images until you see some good, representative examples of things you might expect in operation. Then test out different prompts on the left side of the screen to ensure your model understands your object. For this example, we might experiment with terms like “glasses” or “eyewear.” The model will place bounding boxes where it thinks it sees your object prompt.

If your model adds too many or not enough object identification boxes within an image, you can use the bar in the upper right to change the confidence threshold. A higher confidence threshold directs the model to include only its best guesses. A lower confidence threshold will allow for multiple, lower confidence guesses. The confidence threshold can be changed again later in the model training process.

When you feel good about your prompt and your confidence threshold, click I’m ready to auto label. This step may take a while. The first time around your model is using EyePop.ai’s generic auto-labeler. Each time you work through the human review process and iterate with new data, your own pre-trained model will be used for this step, so it will improve each time you work through the dataset.