Real World Telecom Infrastructure Object Detection

Telkomsel Internship Project – Business Growth & Analytics Division

An applied machine learning project developed during an internship at Telkomsel
to support competitive landscape analysis through visual detection of telecom infrastructure in real-world environments.

---

Project Context

Internship Project

Role : Machine Learning Intern
Division : Business Growth & Analytics
Organization: Telkomsel

This repository represents a technical implementation and experimentation record. It does not contain confidential business logic, internal metrics, or strategic decisions.

---

Business Motivation

Understanding competitor infrastructure presence at the regional level is critical for:

• Market penetration analysis
• Competitive strength estimation
• Infrastructure expansion prioritization
• Field validation support

Manual surveys are slow and costly.

This project explores whether computer vision–based object detection can assist analysts in identifying and quantifying telecom provider infrastructure from street-level imagery as an auxiliary data source for business analysis and preserving a prototype for further improvement and use

---

Technical Objective

Build and evaluate an object detection pipeline capable of:

• Detecting telecom infrastructure objects associated with different providers
• Handling real-world constraints:
  • Severe class imbalance
  • Occlusion and cluttered street scenes
  • Thin and small-scale objects (cables, poles)
• Producing interpretable outputs suitable for downstream analytics, not just model scores

---

Project Status

Functional MVP & Analytic Integration

• Training Pipeline: Refactored from experimental notebooks into modular, reproducible .py scripts.
• Inference Engine: Fully operational with automated EXIF metadata extraction.
• Geo-Spatial Integration: Implemented custom reverse-geocoding to map detection locations to Kecamatan (Sub-district) utilizing BPS (Statistics Indonesia) SHP files.
• Known Limitation: While the pipeline is functional, current model performance may vary under extreme lighting conditions or severe occlusion. Human-in-the-loop verification is recommended for critical decision-making.

---

Tools & Stack

• Model : YOLOv8 (Ultralytics)
• Annotation : CVAT
• Conversion : Datumaro (CVAT → COCO → YOLO)
• Language : Python 3.10+
• Hardware : Kaggle / Colab GPU (T4)

---

Repository Structure

.
├── data/                   # Dataset management & versioning
│   ├── raw/                # Original collected images
│   ├── processed/          # YOLO-formatted datasets ready for training
│   ├── synthetic/          # Generated synthetic data for augmentation
│   └── data_summary.md     # Documentation on class balance & constraints
│
├── deployment/             # Standalone Inference Application
│   ├── map/                # BPS Shapefiles (.shp) for Reverse Geocoding
│   ├── app.py              # Main entry point for the detection tool
│   ├── best.pt             # Optimized Production Model Weights
│   └── requirements.txt    # Dependencies specific for deployment
│
├── experiments/            # R&D Archives
│   ├── notebooks/          # Exploratory analysis & prototyping
│   └── experiments_log.md  # Chronological log of model iterations
│
├── scripts/                # MLOps Pipelines
│   ├── train.py            # Refactored, reproducible training script
│   └── data/               # Data preprocessing & conversion utilities
│
└── README.md

---

Dataset Notes

• Street-level imagery collected from real-world environments
• Provider-level object detection (5 classes)
• Significant real-world class imbalance
• Synthetic data used only to augment training
• Validation and evaluation performed on real data only
• Details are documented in data_summary.md

---

Experiments & Findings

All model iterations, assumptions, failures, and improvements are transparently documented in: experiment_logs.md

This includes:

• Baseline vs augmented training
• Stratified validation strategy
• Synthetic data impact analysis
• Resolution scaling effects
• Known statistical limitations on minority classes

---

Deployment & Geo-Analytics Module

The deployment/ directory contains a self-sufficient tool designed for analysts to process field data locally.

Features:

• Offline Mapping: Uses vector maps stored in deployment/map/ to overlay detection points with administrative boundaries (Kecamatan/Kabupaten).
• Automated Reporting:
  1. Reads images from source folder.
  2. Detects objects using best.pt.
  3. Extracts EXIF Metadata (Lat/Long).
  4. Generates a consolidated CSV report.

How to Run:

# Install deployment dependencies
pip install -r deployment/requirements.txt

# Run the analyzer
python deployment/app.py

---

Intended Use

• Internal experimentation and technical validation
• Supporting analysis for competitive infrastructure mapping
• Demonstration of applied ML engineering during internship

Not Intended For

• Production deployment without further validation
• Automated decision-making without human review
• Disclosure of internal Telkomsel strategy or metrics

Disclaimer

This project is part of an internship assignment. All interpretations and technical decisions represent the author's implementation and do not constitute official Telkomsel products or statements.