Semantic Segmentation PyTorch Practice: Fully Convolutional Network (FCN) and Deconvelutional Network (DeconvNet)

Pierre Jobic, Corentin Barloy and Kexin Ren

This is a Pytorch pratice for the course Deep Learning Do It Yourself 2018/2019 at ENS Paris

Run the model

train.py --dataset_year --model --data --log --epochs --batch --load --check_every --save_every

Description

Data:

PASCAL VOC 2012 segmentation dataset (train + val)

• training images = 1444
• val images = 1464

Data Preprocessing:

color map (3 * 224 * 224) --> one-hot class map (21 * 224 * 224)

FCN training:

• Device: Google Colab w/ GPU
• #epochs = 50; 100
• Batch size = 64
• Learning rate = 0.0001
• Scheduler_lr = (factor = 0.1, each 15 steps)
• Loss func = CrossEntropyLoss
• Pre-trained VGG (ILSVRC dataset)

FCN results:

in results folder, there are 3 versions of FCN -

• v1: lr = 0.0001, 50 epochs
• v2: lr_scheduler(0.1, 15) (initial lr = 0.0001), 50 epochs
• v3: lr_scheduler(0.1, 20) (initial lr = 0.0001), 100 epochs

DeconvNet training:

• Device: Google Colab w/ GPU
• #epochs = 50
• Batch size = 64
• Learning rate = 0.001
• Loss func = MSELoss

DeconvNet results:

can be found in results/First Test folder

Reference:

FCN: Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 3431-3440).

DeconvNet: Noh, H., Hong, S., & Han, B. (2015). Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE international conference on computer vision (pp. 1520-1528).