Semi-Supervised Training with One-stage Object Detection

Abstract: Training a specialized object detection model requires large amounts of labeled data that may not be easily obtainable. In contrast, large amounts of unlabeled data relevant to the task are often available. In this report a semi-supervised approach that increases the amount of training data by adding soft labels is examined. Soft labels are generated during training by filtering predictions made on the unlabeled dataset. This training approach was used with CenterNet, a one-stage detection model based on predicting keypoints of objects in an image. Several loss functions incorporating the soft labels are tested. A minor improvement in detecting harder classes was achieved, however semi-supervised learning did not significantly improve CenterNet’s overall performance.