The paper introduces a method for predicting the performance of convolutional long short-term memory networks, particularly in tasks that require semantic segmentation. The method involves estimating predictive quality based on temporal cell state-based input metrics per segment. The authors also explore the effect of the number of cell states considered for the proposed metrics. The study is significant in safety-critical applications like autonomous driving where video data is available, and reliable prediction is indispensable. The method achieves a classification accuracy of 96.15% and AUROC of 95.04%.
Publication date: 13 Nov 2023
Project Page: https://arxiv.org/abs/2311.07477v1
Paper: https://arxiv.org/pdf/2311.07477
