[深度学习论文笔记][Semantic Segmentation] Recurrent Convolutional Neural Networks for Scene Labeling

Pinheiro, Pedro H da Silva, and Ronan Combes. "This paper presents Recurrent Convolutional Neural Networks for Scene Labeling." This paper appeared in the proceedings of ICML in 2014. (Referenced in 163 papers).

1 Pipeline

See Fig. Each instance accepts input comprising a resized RGB image and the classification predictions from its preceding network stage. The initial instance's classification predictions are initialized as zero maps.

2 Training Details

The network undergoes end-to-end training, meaning its components are integrated and trained in a unified process without separation. The weights across all instances are shared, indicating that each instance contributes equally to a common set of parameters. The total loss incurred by the entire structure equates to a summation of individual CNN instance losses, ensuring that each component's contribution to overall performance is accurately reflected.

3 Results

Check Figure. A number of recurrent iterations yield better outcomes. This is primarily due to a number of recurrent iterations enabling the model to simultaneously consider a wider range of context without causing its capacity to become overly constrained.

model, and iteratively refine the previous predictions.