ndilsou / robond-fcn-segmentation Goto Github PK

The objective of this project was to train a fully convolutional network to perform semantic segmentation and allow for target tracking on a quadcopter. Final average IoU obtained after 22 epochs of training is 0.475.

All the files and their content is described below:

• writeup.md: Notebook containing the network architecture and used for training.

• code/model_training.ipnb: Notebook containing the network architecture and used for training.

• code/data_augmentation.ipnb: Contains code used to perform horizontal flipping on the images and double the dataset size.

• data/weights: Weights of the trained networks.

• code/keras_viz_dependencies.txt: Requirement for keras.utils.vis_utils.plot_model.

• docs/misc/model_architecture.png: Picture of the network architecture.

We will define blocks using the following nomenclature: BLOCK(kernelWxkernelH, Stride, depth) when relevant. Our network is composed of an assembly of 5 main building blocks:

• encoder blocks : ENCODER(depth)=[ SEPARABLE_CONV2D(3x3, 1) -> RELU -> BATCHNORM ]

• decoder blocks : DECODER(concat, depth)=[ BILINEAR_UPSAMPLE(2x2) -> CONCATENATE(concat) -> [ SEPARABLE_CONV2D(3x3, 1) -> RELU -> BATCHNORM ] * 2 ]

• downsampling blocks : DOWNSAMPLE=[MAXPOOL(2x2, 2)] OR [ENCODER(2x2, 2)]

• dropout blocks : DROP=[SPATIAL_DROPOUT2D]

• 1x1 convolutional block : 1x1CONV2D(depth)=[CONV2D(1x1, 1) -> RELU -> BATCHNORM]

The picture of the final architecture is available here. This architecture is inspired from SegNet.

The full architecture of the final model using the previous notation is as follow:

[INPUT] -> [ENCODER(32) -> DOWNSAMPLE -> DROP] -> [ENCODER(64) -> DOWNSAMPLE -> DROP] -> [ENCODER(128) -> DOWNSAMPLE -> DROP] -> [1x1CONV2D(256)] -> [DECODER(encoder128, 128) -> DROP] -> [DECODER(encoder64, 64) -> DROP] -> [DECODER(encoder32, 32) -> DROP] -> [OUTPUT]

Note that we used MAXPOOL for all the final model downsampling.

Total params: 142,814
Trainable params: 140,958
Non-trainable params: 1,856

Evaluation Set size:

Average IoU:

Confusion Table:

Example predictions:

left: input image, middle: target mask, right: output mask.