Vision for Pollution - Deep Learning to detect visual pollution

View the full paper → Supervisors: Karthik Mohan and Sohan Seth

Final Grade: 82%

Conference submission: This project was later adapted for a submission to the British Machine Vision Conference (BMVC). View →

Overview

Visual pollution is caused by man-made elements such as advertisements, infrastructure, and urban clutter that negatively affect perceived environmental quality and human wellbeing.

This project presents a scalable computer vision framework for detecting and quantifying visual pollution using deep learning and large-scale street-view imagery.

Methods

The project followed a large-scale data collection and analysis pipeline:

1. City selection
   Used a public global cities dataset to identify cities with populations greater than 100,000.

2. Street-view data collection
   Sourced geolocated street-view imagery from Mapillary within each city’s boundary and stored the associated metadata.

3. Object detection
   Trained a YOLO26 model to detect nine classes of visual pollution.

   Billboards, Utility Poles, Graffiti, Mobile Advertisement, Shop Signs, Road Signs, Barriers, Potholes

4. Large-scale inference
   Ran the trained model on more than 2,000,000 images to measure visual pollution across cities worldwide.

5. Visual Pollution Index
   Developed a Visual Pollution Index (VPI) to quantify and compare pollution levels between regions.

Results

Visual Pollution Index

The VPI was used to rank cities according to their measured levels of visual pollution.

Highest Pollution City	Country	VPI	Lowest Pollution City	Country	VPI
Kolkata	India	0.653	Shenzhen	China	0.071
Pune	India	0.568	Lahti	Finland	0.078
Barman Kalan	India	0.567	Mission Viejo	United States	0.082
Dhaka	Bangladesh	0.563	Siracusa	Italy	0.086
Kumasi	Ghana	0.562	Columbia	United States	0.089
Caloocan City	Philippines	0.560	Roseville	United States	0.099
Malang	Indonesia	0.538	Messina	Italy	0.110
Chennai	India	0.532	Pomona	United States	0.112
Purwokerto	Indonesia	0.531	Stockholm	Sweden	0.112
Vishakhapatnam	India	0.523	Serpukhov	Russia	0.113

Global VPI Map

The resulting scores were used to construct a global map showing the geographic distribution of visual pollution across the evaluated cities. It shows that while western countries have more cities with lower VPI scores, there are also more cities with sufficient data to calculate a score.

Poland Advertising Regulation Experiment

In 2015, Poland introduced legislation giving municipalities greater powers to regulate the aesthetics of public spaces, including advertising and signage.

To investigate whether the proposed framework could detect changes over time, billboard prevalence was compared before and after local regulations were introduced in three Polish cities.

City	Period	Images	Billboards	Density	Change
Kraków	2005–2019	3,600	876	0.243	—
	2020–2026	7,941	967	0.122	−49.9%
Gdańsk	2005–2017	1,901	550	0.289	—
	2018–2026	2,059	353	0.171	−40.8%
Wrocław	2005–2019	4,005	1,035	0.258	—
	2020–2026	4,295	680	0.158	−38.7%

Billboard density declined substantially across all three cities, with reductions ranging from 38.7% to 49.9%.

These results do not establish that the regulations directly caused the decline. However, they demonstrate that the proposed framework can detect and quantify long-term changes in visual pollution using historical street-view imagery.

Limitations

Data Availability

The primary limitation was the availability of street-view imagery. Since the project relied on crowd-sourced data, coverage varied considerably between countries and cities.

Wealthier and more frequently mapped regions generally contained substantially more imagery, producing more reliable VPI estimates. Cities with limited coverage could not always be evaluated with the same level of confidence.

Training Data Imbalance

The second major limitation was variation in the amount of training data available for different pollutant classes.

Common visual pollutants, such as billboards, appeared frequently in available imagery and therefore had more training examples. Less common or harder-to-identify pollutants, such as potholes, had fewer examples and consequently produced less reliable detections.