Object Recognition under Lighting Variations using Pre-Trained Networks

We report the object-recognition performance of VGG16, ResNet, and SqueezeNet, three state-of-the-art Convolutional Neural Networks (CNNs) trained on ImageNet, across 15 different lighting conditions using the Phos dataset and a ResNet-like network trained on Pascal VOC on the ExDark dataset. The instabilities in the normalized softmax values are used to highlight that pre-trained networks are not robust to lighting variations. Our investigation yields a robustness analysis framework for analyzing the performance of CNNs under different lighting conditions.The Phos dataset consists of 15 scenes captured under different illumination conditions: 9 images captured under various strengths of uniform illumination, and 6 images under different degrees of non-uniform illumination. The ExDARK dataset consists of ten scenes under different illumination conditions. A Keras-based pipeline was developed to study the softmax values output by ImageNet-trained VGG16, ResNet, and SqueezeNet for the same object under the 15 different lighting conditions of the Phos dataset. A ResNet architecture was trained end-to-end on the PASCAL VOC dataset. Large variations observed in the softmax values provide empirical evidence of unstable performance and the need to augment training to account for lighting variations.