Pinheiro, Pedro HO, and Ronan Collobert. “Recurrent Convolutional Neural Networks for Scene Labeling.” ICML. 2014. (Citations: 163).
1 Pipeline
See Fig. Each instance takes as input both an resize RGB image and the classification predictions of the previous instance of the network. The classification predictions of the first instance are zero maps.
2 Training Details
The network is trained end-to-end. The weights of all the instances are the shared. The loss the whole structure is the sum of the losses of each CNN instance.
3 Results
See Fig. More recurrent iterations improve results. This is mainly because more iterations can let the model consider large receptive fields while limiting the capacity of themodel, and iteratively refine the previous predictions.