Exploring Image Similarity-Based Splitting Techniques in Automotive Perception Systems

Training object detection models for automotive perception systems is often a challenging task due to the variation in the image data used to train and test the model. The images might come from different geographic locations as well as different weather and lighting conditions. Image similarity-based split could be a safe option to keep the train and test sets as identical as possible. However, based on which similarity measure is chosen to split the data, the train and test set may not contain enough representative images of all the different situations, including seasonal and lighting variations, and hence often degrades overall performance. This study considered four different image similarity measures based on visual features and intrinsic/semantic information that come with the image data. The semantic similarity helps to avoid gathering images in the same (train/test) set that visually look similar to each other. The evaluation results show that the semantic similarity-based splits resulted in 12–47% higher performance of the object detection model in terms of mean average precision (mAP) and F1-score. Among the four similarity measures, AllClass similarity consists of the highest intrinsic information available with the image data, which also results in the highest performance of the model when used for data splitting.