TECHFEST IIT BOMBAY

HACKATHON

Problem Statement:Object Detection Model

Team Slot : 11

Team Members: Aaryan Sahu

Abhinav Singh Bhaduria
Suryansh Sharma
Problem Statement:Develop an object detection model that can accurately detect, classify, and
localize multiple objects within an image or video frame.

Proposed Solution: Objectives:

• For solving the problem statement we are using a multi • detect objects with high precision,
modal approach in we can combine the results from more minimizing false positives and false
than one model and based on these predicted result and by negatives.
their mean we will predict the final result. • process images or video in real-time or
• We have used mahcine learning and deep learning models near real-time.
like • practical for real-world applications.
YOLO,EfficientDet for faster results and more accuracy. • should handle diverse environments
• We have used this pre-trained model and fine tuned them
on dataset COCO and created our modal which have
accuracy combined of above and can predict the results Uniqueness:
precisely. • Multi modal Approach
• Our product will be a web based application which can • Real Time Detection
be utilised by anyone so that they can use it for real • Real-Time Object Detection with Webcam,
time monitoring or can for detection in images and • Custom Alerts
videos. • Compatible With CCTV, and External
Camera.
Scalability: TechStack:
Frontend- HTML,CSS,JS,Bootstrap.
• We are planning to upgrade the model to real time Backend-Python(Flask)
monitoring so that the live videos like of CCTV’s Python Libraries- Opencv,TensorFlow,Keras
can also be analyzed. Pre-Trained Model- Ultralytics,YOLO,
• We can create our custom dataset and fine tune Efficientdet
the model on it so that it can get to know about Data Handling Library-Pandas,Numpy
Dataset-COCO
more classes of object detection.
• Fine Detections: Differentiate between objects of Technical Challenges:
same class like different species of birds and 1. Computational Limitations
dogs. High GPU Memory Requirement
Solution: Can use cloud notebooks like Kaggle,Google
• Can be used in Smart Vehices for calculating the Colab.
distance of upcoming objects 2 Model Challenges
Problem: The model performs well on training data but
poorly on new images.
Solution:
Apply techniques like dropout, weight regularization, and
data augmentation.
3. Dataset Challenges
Problem: Certain classes in the COCO dataset may have
fewer instances, leading to biased predictions.
Solution: Use class re-weighting or data augmentation
techniques to balance the dataset.
 Real World Applications
Technical Feasibility:
Proven Effectiveness: YOLO (You Only Look Once) is a
Automation of Tasks: state-of-the-art model with a strong track record in object
Enables machines to identify and locate objects in detection tasks.
images or videos, reducing the need for manual Pre-trained Datasets: Using COCO reduces training
intervention in tasks like surveillance, sorting, and complexity and ensures accurate results for a wide range of
monitoring. objects.
Real-Time Applications: Real-Time Compatibility: YOLO’s efficiency makes it suitable
The YOLO model's speed and accuracy make it for real-time applications on moderately powered GPUs.
suitable for real-time object detection in areas like Python Libraries: The project leverages mature libraries like
autonomous vehicles, security systems, and robotics. PyTorch, OpenCV, and Flask for seamless implementation
Adaptability Across Domains: and deployment.
Fine-tuning on the COCO dataset allows for diverse Viability:
object detection applications, including retail inventory Cost Effectiveness
management, wildlife monitoring, and traffic control. Development Costs: Free tools (YOLO, Python libraries) and
Cost-Effectiveness: datasets minimize the cost.
Automation reduces human labor and minimizes errors, Deployment Costs: Cloud platforms can be used temporarily;
saving operational costs in industries like logistics and edge deployment options are cost-efficient for long-term use.
manufacturing. Scalability: Practical Applications
The model can be fine-tuned further or extended to detect Diverse Use Cases: Applications in security (surveillance),
additional objects, making it adaptable to new datasets retail (inventory management), and healthcare (medical
and use cases. imaging).
Customizability: Fine-tuning allows the model to adapt to
specific requirements like detecting niche objects.
 UserInterface

Core Features

• Image Detection

• Webcam Detection

• Video Detection

• Custom Alerts
 Detection Results
For an input video our model povides with a table of list of different
objects detected and also gives an output video with bounding
boxes of different labels.

Classes Included:
• Persons
• Cat
• Dogs
• Car
• Truck
• Cell Phone
• TV

90+ Classes
THANK YOU