Developing efficient transfer learning strategies for robust scene recognition in mobile robotics using pre-trained convolutional neural networks
arXiv:2107.11187
Abstract
We present four different robust transfer learning and data augmentation strategies for robust mobile scene recognition. By training three mobile-ready (EfficientNetB0, MobileNetV2, MobileNetV3) and two large-scale baseline (VGG16, ResNet50) convolutional neural network architectures on the widely available Event8, Scene15, Stanford40, and MIT67 datasets, we show the generalization ability of our transfer learning strategies. Furthermore, we tested the robustness of our transfer learning strategies under viewpoint and lighting changes using the KTH-Idol2 database. Also, the impact of inference optimization techniques on the general performance and the robustness under different transfer learning strategies is evaluated. Experimental results show that when employing transfer learning, Fine-Tuning in combination with extensive data augmentation improves the general accuracy and robustness in mobile scene recognition. We achieved state-of-the-art results using various baseline convolutional neural networks and showed the robustness against lighting and viewpoint changes in challenging mobile robot place recognition.
arXiv comment18 pages, 1 figures, 10 tables. Submitted to IEEE Transactions on Robotics (T-RO)
References in corpus (11)
- MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
- VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator
- DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition
- How transferable are features in deep neural networks?
- MobileNetV2: Inverted Residuals and Linear Bottlenecks
- Do Deep Nets Really Need to be Deep?
- On the Performance of ConvNet Features for Place Recognition
- A Discriminative Representation of Convolutional Features for Indoor Scene Recognition
- Up or Down? Adaptive Rounding for Post-Training Quantization
- Empty Cities: a Dynamic-Object-Invariant Space for Visual SLAM
- Learning Deep NBNN Representations for Robust Place Categorization