The effect of instance mask-guided attention for discriminative learning of vehicular collision image classification

A vehicular collision is an unfortunate event and a great challenge to be addressed for autonomous vehicles’ navigation. In this paper, an instance mask-guided attention (IMGA) method is presented for vehicular collision image classification. The major challenge in developing deep learning solutions to understand vehicular collision lies in the dataset, which is typically developed from unintentionally captured dashboard camera videos that have wide variations in acquisition conditions. In the proposed IMGA method, the instance masks of vehicles undergoing collision in an image supervise the learning of an attention-based image classification model. The proposed method consists of a pre-trained instance segmentation branch that generates colliding vehicles instance masks that are used as ground truth for learning the attention map of the classifier. This helps achieve a discriminative learning strategy that concentrates only on the significant regions in the image from its background increasing its robustness and accuracy. Experimentation proved the effectiveness of the proposed IMGA method with a result of 97.94% accuracy in the presented vehicular collision image classification dataset which is much higher compared to state-of-the-art image classification models.