U-Net Implementation in TensorFlow

Re implementation of U-Net in Tensorflow

• to check how image segmentations can be used for detection problems

Original Paper

• U-Net: Convolutional Networks for Biomedical Image Segmentation

Summary

Vehicle Detection using U-Net

Objective: detect vehicles Find a function f such that y = f(X)

Input	Shape	Explanation	Example
X: 3-D Tensor	(640, 960, 3)	RGB image in an array
y: 3-D Tensor	(640, 960, 1)	Binarized image. Bacground is 0 vehicle is masked as 255

Loss function: maximize IOU

    (intersection of prediction & grount truth)
    -------------------------------------------
    (union of prediction & ground truth)

Examples on Test Data: trained for 3 epochs

Get Started

Download dataset

• the annotated driving dataset is provided by Udacity
• In total, 9,423 frames with 65,000 labels at 1920x1200 resolution.

make download

Resize image and generate mask images

• utils/data.py is used to resize images and generate masks

make generate

Train Test Split

Make sure masks and bounding boxes

jupyter notebook "Visualization & Train Test Split.ipynb"

Train

# Train for 1 epoch
python train.py

or

$ python train.py --help
usage: train.py [-h] [--epochs EPOCHS] [--batch-size BATCH_SIZE]
                [--logdir LOGDIR] [--reg REG] [--ckdir CKDIR]

optional arguments:
  -h, --help            show this help message and exit
  --epochs EPOCHS       Number of epochs (default: 1)
  --batch-size BATCH_SIZE
                        Batch size (default: 4)
  --logdir LOGDIR       Tensorboard log directory (default: logdir)
  --reg REG             L2 Regularizer Term (default: 0.1)
  --ckdir CKDIR         Checkpoint directory (default: models)

Test

• Open the Jupyter notebook file to run against test data

jupyter notebook "./Test Run After Training.ipynb"