Enhancing Detection Accuracy with Fine-Tuning on BrightEarth

With BrightEarth, you may encounter missing elements or false positives (hallucinations). To address this, we offer a fine-tuning function that improves detection accuracy for your uploaded images. By retraining the model on a defined area with your ground truth data, it becomes more robust, leading to better extractions across the image. Here’s how it works:

watch this video on youtube

Step 1: Prepare your ground truth data

First, ensure you have an orthorectified source image, in which you also have a set of highly reliable, verified vector data for a small, specific area correlated with your image. This data serves as the basis for fine-tuning.

Step 2: Upload ground truth data

Create a project whose area of interest includes the whole of your image. Load the ground truth dataset into this project. This consists of the imagery, the tagged geometry shapefile, and the area of interest delimiting the ground truth data.

Step 3: Retrain the model

Set up the fine-tuning process by configuring a scenario. Select the appropriate dataset and specify:

•   The source image
•   The ground truth area of interest
•   The labeled geometries
•   The type of detector you wish to enhance

 Then click on Execute to retrain the models.

Step 4: Run a new extraction

Once the model has been refined, run a new extraction in your project, selecting the newly refined model. The model now applies to the entire image, and its improved understanding of your area of interest will give you better results across your entire image.

Step 5: Compare results

Finally, the results obtained before and after fine-tuning can be compared. You’ll see a significant improvement in detection accuracy and a reduction in errors or false positives.

This new and improved detector can also be used on other images with similar characteristics.