HOG-supervised compact CNNs for real-time visual place recognition

This paper presents HOGNet, an efficient deep learning framework aimed at replicating and improving histogram of oriented gradients (HOG) features for visual place recognition (VPR). HOGNet distinguishes itself from traditional deep models by utilizing handcrafted descriptors for efficiency and interpretability, while simultaneously harnessing the representational capabilities of convolutional neural networks (CNNs), thus reducing the need for extensive labeled datasets and significant computational resources. Two variants of the model, HOGNet_441 and HOGNet_1764, are developed to predict HOG descriptors of varying lengths, facilitating adaptable implementation in diverse computational environments. We assess HOGNet using two challenging datasets, the Nordland dataset and the India driving dataset (IDD-VPR), and perform thorough comparisons with established architectures such as VGG16, ResNet101, and MobileNet. HOGNet_441 achieves 67.7% validation accuracy and 90.76% Recall@20, outperforming VGG16 (88.00% at Recall@20) while being 36x faster (3.44 ms vs. 124.17 ms). An ablation study confirms that HOG supervision provides a +5.55% accuracy improvement over a baseline CNN without HOG targets. Intermediate convolutional layers (Pool2, Pool3) are shown to be more effective for VPR than output layers. HOGNet achieves approximately 290 FPS, making it suitable for real-time and resource-constrained applications. The findings underscore the potential of incorporating handcrafted feature descriptors into deep learning models to attain an optimal balance among accuracy, interpretability, and computational efficiency.