CV25 Intermediate Project – Object Detection

Table of Contents

• Build Instructions
• Code Design and Style Guide
• Objective
  • Inputs
  • Outputs
  • Suggested Approach
  • Dataset
  • Evaluation Metrics
  • Project Requirements
  • Submission Checklist

Build Instructions

# To start the container on the VLAB env
start_opencv
# Create the build directory in the root of the project 
mkdir build
cd build

# Configure the project using CMake
cmake ..

# Compile the project
make

# Run the executable
./CV_Midterm

The above instructions will create a /results directory with two subdirectories one for the /images and the other for the /coordinates of the bounding boxs computed by the algorithm.

Code Design and Style Guide

All contributors swear to follow these points with the utmost care:

• No usage of GLOBAL VARIABLES!
• Final project musn't have binary files included
• Usage of a single CMakeLists.txt with more add_executable commands
• Compile the code, and make sure it runs bug free before pushing
• Include guards to be added in Header file
• A header file musn't contain function definitions, only declerations!
• Never #include a .cpp file
• Never add absolute paths in code
• Always check command-line arguments, before processing them, use std::cerr for debugging purposes
• Define GUI names as const variables
• Don't mix more than one activity in a single function, for example maxfilter should only apply the filter, not visualize it as well
• Pass by reference or pointer to be used in functions, for example const cv::Mat& src
• For the sake of clarity:
  • No usage of namespaces
  • Only camelCase to be used, fyi case guide
  • When it comes to comments, follow only the the google c++ style guide

Objective

Develop an object detection system using C++ with OpenCV to identify and localize three known objects:

• Power Drill
• Mustard Bottle
• Sugar Box

The system should detect each object in a scene image and draw a bounding box around it.

---

Inputs

• RGB scene image
• Synthetic views of each object
• Binary segmentation masks (optional)

---

Outputs

• Text file containing bounding box coordinates
• Scene image with bounding boxes overlaid

---

Suggested Approach

Use robust feature matching as the base method. Other allowed methods:

• Template matching
• Appearance-based detection
• Bag of Words (BoW)

Deep learning methods are not allowed.

---

Dataset

Dataset Download

Each object has:

• models/ — 60 synthetic views + segmentation masks
• test_images/ — Images with scenes for detection
• labels/ — Ground truth in the format: <object_id>_<object_name>

---

Evaluation Metrics

• Mean Intersection over Union (mIoU): Average IoU across all object categories
• Detection Accuracy: Number of correct detections (IoU > 0.5 with ground truth)

---

Project Requirements

• Implemented in C++ with OpenCV
• Must compile and run in the official Virtual Lab
• Only original code written by group members is allowed

---

Submission Checklist

• Source code (C++), one author per file
• Report (max 2 pages of text, excluding images/tables), including:
• Approach explanation
• Performance results (metrics + images)
• Team member contributions
• Hours worked per person

---