The paper presents a method to improve object detection in infrared (IR) images by leveraging cues from the RGB modality. The lack of sufficient labeled training data in the IR modality is identified as a major challenge. The authors propose a novel tensor decomposition method, called TensorFact, which splits the convolution kernels of a layer of a Convolutional Neural Network (CNN) into low-rank factor matrices. These matrices are pre-trained on the RGB modality and then augmented with a few trainable parameters for training on the IR modality. The approach is validated by assessing its performance on both RGB and IR images. The findings suggest that TensorFact shows performance gains on RGB images and outperforms a standard state-of-the-art object detector on the FLIR ADAS v1 dataset by about 4% in terms of mAP 50 score.
Publication date: 28 Sep 2023
Project Page: https://arxiv.org/abs/2309.16592
Paper: https://arxiv.org/pdf/2309.16592
