Human Centric Visual Analysis with Deep Learning by Liang Lin & Dongyu Zhang & Ping Luo & Wangmeng Zuo

Author:Liang Lin & Dongyu Zhang & Ping Luo & Wangmeng Zuo
Language: eng
Format: epub
ISBN: 9789811323874
Publisher: Springer Singapore

(6.1)

The resolution of the output score maps is , which is the same for both segmentation and edge. Thus, the segmentation branch has a -dimensional output, which encodes K segmentation maps of resolution , one for each of the K classes. During training, we apply a per-pixel softmax and define as the multinomial cross-entropy loss. is the same but for the refined segmentation results. For each -dimensional edge output, we use a per-pixel sigmoid binary cross-entropy loss. , , and denote the loss of the first predicted edge, refined edge, and side-output edge, respectively. In our network, the number of edge side outputs, N, is 3. and are the balance weights.

We use the batch normalization parameters provided by [10], which are fixed during our training process. Our modules (including the ASPP and pyramid pooling module) added to ResNet eliminate batch normalization because the whole network is trained end-to-end with a small batch size due to the limitation of physical memory on GPU cards. The ReLU activation function is applied following each convolutional layer except the final classification layers.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.