Batch regularization to converge the deep neural network for indoor RGBD scene understanding
Hassan Hayat, Yazhou Liu, Maqsood Hussain Shah, Adnan Ahmad · 2017
Training a deep neural network is complicated due to the input distribution of each layer changes during training. Small changes are amplified throughout the network and consequently the covariate-shift is likely to occur. That is why small learning is critical but small learning rates is the root to slow training process and may even prevent the escape of suboptimal local minima. This paper purposed a normalization step before fusing the classification scores of different strides not the feature vectors. This emphasizes the influence of the coarsest classifier in the beginning. The influences of the finer predictions can adapt gradually. Experiment results have shown that this technique assists the learning process and leads to better convergence for training deep neural network and presents the better understanding of indoor RGBD scenes.