README.md

Building Footprint Segmentation

Library to train building footprint on satellite and aerial imagery.

Installation

pip install building-footprint-segmentation

Dataset

• Massachusetts Buildings Dataset
• Inria Aerial Image Labeling Dataset

Training

• Train With Config , Use config template for generating training config

• Train With Arguments

Visualize Training

Test images at end of every epoch

• Follow Example

Visualizing on Tensorboard

from building_footprint_segmentation.helpers.callbacks import CallbackList, TensorBoardCallback
where_to_log_the_callback = r"path_to_log_callback"   
callbacks = CallbackList()

# Ouptut from all the callbacks caller will be stored at the path specified in log_dir
callbacks.append(TensorBoardCallback(where_to_log_the_callback))

To view Tensorboard dash board

tensorboard --logdir="path_to_log_callback"

Defining Custom Callback

from building_footprint_segmentation.helpers.callbacks import CallbackList, Callback

class CustomCallback(Callback):
    def __init__(self, log_dir):
        super().__init__(log_dir)


where_to_log_the_callback = r"path_to_log_callback"   
callbacks = CallbackList()

# Ouptut from all the callbacks caller will be stored at the path specified in log_dir
callbacks.append(CustomCallback(where_to_log_the_callback))

Split the images in smaller sample

import glob
import os

from image_fragment.fragment import ImageFragment

# FOR .jpg, .png, .jpeg
from imageio import imread, imsave

# FOR .tiff
from tifffile import imread, imsave

ORIGINAL_DIM_OF_IMAGE = (1500, 1500, 3)
CROP_TO_DIM = (384, 384, 3)

image_fragment = ImageFragment.image_fragment_3d(
    fragment_size=(384, 384, 3), org_size=ORIGINAL_DIM_OF_IMAGE
)

IMAGE_DIR = r"pth\to\input\dir"
SAVE_DIR = r"pth\to\save\dir"

for file in glob.glob(
    os.path.join(IMAGE_DIR, "*")
):
    image = imread(file)
    for i, fragment in enumerate(image_fragment):
        # GET DATA THAT BELONGS TO THE FRAGMENT
        fragmented_image = fragment.get_fragment_data(image)

        imsave(
            os.path.join(
                SAVE_DIR,
                f"{i}_{os.path.basename(file)}",
            ),
            fragmented_image,
        )

Inference

• Perform Augmentation during Inference and aggregate results

Segmentation for building footprint

• binary
• building with boundary (multi class segmentation)

Weight File

• RefineNet trained on INRIA
• DlinkNet trained on Massachusetts Buildings Dataset

Refer gtkit for commonly used utility task when working with Geotiff

• Generate bitmap from shape file
• Generate shape geometry from geo reference bitmap
• Save Multi Band Imagery