Building Classification

This is the code for the project of soft-story building classification.

The goal of this project is designing a model which can automatically classify a soft-story building based on a single street view image.

Prerequisites

Tensorflow 1.0.

Python 2

CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

step 1: Install latest version of TF-slim following the instruction here
step 2: Put this repo in the foloder /models/research/slim/
step 3: Download pre-trained models (ResNet50/152 or InceptionV3/V4) and put them in the folder /models/research/slim/pretrained/; download data (Santa Monica and Oakland and put them in /models/research/slim/tfrecords/
step 4: Run the code.

Training a Model

  cd /models/research/slim/
  ./finetune_resnet_50_on_buildings.sh

Evaluating a Model

Please download our trained models here (923M).

  cd /models/research/slim/
  ./finetune_resnet_50_on_buildings_eval.sh

Datasets and Results

Datasets (SS refers to soft-story building). All images in the datasets are collected by Google Street View API.

City	# SS	# non-SS	# train	# test
Santa Monica	3,203	3,921	6,421	712
Oakland	717	642	1,224	135

Performance of ResNet50 / InceptionV3 on Santa Monica / Oakland

Model	avg. acc.	Precision	Recall	F1
ResNet50	85.94%	84.16%	82.80%	83.47%
InceptionV3	84.38%	81.39%	83.77%	82.56%

Model	avg. acc.	Precision	Recall	F1
ResNet50	82.29%	81.54%	82.81%	82.17%
InceptionV3	80.21%	80.65%	78.13%	79.37%

Generalization ability of the models

Note: The models are trained on Santa Monica dataset and tested on 395 street view images collected from Berkeley and San Jose.

Model	avg. acc.	Precision	Recall	F1
ResNet50	86.61%	84.26%	89.34%	86.70%
InceptionV3	87.72%	84.26%	92.39%	88.14%

Application

Given a specific city/region, a soft-story building distribution map can be created based on the prediction of the trained model. The below figure shows the distribution map of Oakland, which is created by SURF.