November 2021 Summaries

This tutorial demonstrates how to use Label Studio, an open-source data labeling tool, and Hub, a dataset format for AI, to simplify semantic segmentation projects. Semantic segmentation involves attributing a class to each pixel of an image, such as identifying whether a pixel belongs to the "smile" or "non-smiling" class in this case. The tutorial uses the GENKI-4K subset containing 4000 face images labeled as smiling or non-smiling by human coders. After filtering for smiling images and labeling smiles using Label Studio, the resulting dataset is saved to Hub Storage. Finally, a semantic segmentation model (UNet) is trained on this data, with results showing how pixels where probabilities exceed a certain threshold are considered "smiling."

This tutorial demonstrates how to simplify semantic segmentation projects using Label Studio and Hub, focusing on identifying smiles in images. Semantic segmentation involves classifying each pixel in an image, and in this case, determining whether pixels belong to a "smiling" or "non-smiling" class. The project utilizes the GENKI-4K dataset, which contains 4000 face images labeled for smile presence. After filtering and selecting images labeled as "smiling," the tutorial guides users through labeling these images using Label Studio and exporting the labeled data. The data is then used to create a computer vision dataset compatible with Hub, facilitating the training of a semantic segmentation model, specifically a UNet model. The training process involves resizing images, ensuring binary mask values, and using TensorFlow to create a dataset for training the model. The tutorial includes instructions for setting up the model, compiling it with metrics like binary accuracy, and executing training with callbacks for optimization. Finally, the tutorial covers testing the model on non-labeled images, adjusting probability thresholds for predictions, and encourages further labeling and parameter optimization for improved results.