Scribd
Upload
23 views23 pages

Demo Day Presentation

The document outlines a project to develop an automatic car classification system using deep learning. It includes 10 sections: structure, deliverables, objectives definition, literature review selecting models, workplan, exploratory data analysis and preprocessing, main model selection (DenseNet), project demo, conclusions, and next steps. The team analyzed vehicle image data, tested models like YOLOv7, and selected DenseNet which achieved 85% test accuracy for classifying make and model of vehicles. Future work includes improving data quality and expanding the model capabilities.

Uploaded by

Sergio Toral Calderon
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
23 views23 pages

Demo Day Presentation

The document outlines a project to develop an automatic car classification system using deep learning. It includes 10 sections: structure, deliverables, objectives definition, literature review selecting models, workplan, exploratory data analysis and preprocessing, main model selection (DenseNet), project demo, conclusions, and next steps. The team analyzed vehicle image data, tested models like YOLOv7, and selected DenseNet which achieved 85% test accuracy for classifying make and model of vehicles. Future work includes improving data quality and expanding the model capabilities.

Uploaded by

Sergio Toral Calderon
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 23

Nahuel Garcia Enzo Gianotti

Team 01
Joheel Luna
Carlos Prado
Automatic Car Classification

Agustin Genou Nicolas Passadore

anyoneai.com
Structure
1. Deliverables
2. Definition of the objective of this work
3. Literature Review for model Selection
4. Workplan
a. Designed Workflow
b. Deployment: Defined Architecture
c. Definition of profiles
5. Exploratory Data Analysis and Preprocessing
a. Analysis
b. Problems found

c. Use of classification models


■ Custom binary Classification Model
■ YoloV7
d. Remove Background
6. Main Model Selected
a. Metrics
7. Project DEMO
8. Conclusions
9. Next Steps
10. Q&A
anyoneai.com
Deliverables
- Vehicle classifier Convolutional Neural Network (CNN) for make and model for
e-commerce
- Exploratory Data Analysis (EDA)
- Pre-processing, data preparation and training scripts
- Application Programming Interface (API) and basic User Interface (UI)
- Containerized with Docker and ready to be deployed
- Unit tests
- OCR

anyoneai.com
1. Objectives Definition

- Online car marketplaces need to


process and organize large
volumes of images in their
websites.
- Users need to reduce the time
they spend on their online bidding
process.
- Cazoo is being forecast to record
a compound annual growth rate
of 209 percent between 2020 and
2022 (Statista).

anyoneai.com
1.a Analytic Approach

- The business problem can be


solved by implementing an image
classification Machine Learning
system that predicts the make
and model of an uploaded image.
- The system deployed has to filter
out corrupted files and
non-vehicle images.

anyoneai.com
1. “A Systematic Evaluation of Recent Deep Learning Architectures for
Fine-Grained Vehicle Classification” by Valev, Sommer and Beyerer

● Objective: compare the performance of Convolutional Neural Network (CNN) architectures for
classification tasks
● Dataset: Stanford Cars-196 dataset.
● Methodology: adapt existing CNN for fine-grained vehicle classification transfer learning and data
augmentation
● Evaluation: compare architectures, approach (from scratch and fine-tuned) and accuracy
● Results: DenseNet-161 had the highest accuracy
● Conclusion: it’s better to use fine-tuned architectures trained on ImageNet with data augmentation

anyoneai.com
2. “Data Augmentation and Clustering for Vehicle Make/Model
Classification” by Nafzi, Brauckmann and Glasmachers

● Objective: improve the accuracy of a vehicle classification model by using clustering


● Dataset: Stanford data set and video data.
● Data Augmentation: web crawler to download data of older car models.
● Clustering: matching score of two feature vectors (client score or impostor score) to
increase classes for each year
● CNN-Architecture: based on Res-Net
● Conclusion: clustering is useful for data augmentation with vehicles with different
released years and/or different perspectives.

anyoneai.com
3. WorkPlan
a. Designed WorkFlow

anyoneai.com
3. WorkPlan
b Designed Architecture

anyoneai.com
3. WorkPlan
c Definition of profiles
● Definition:
○ Objectives
○ Tasks and order of their execution
○ Task compliance monitoring tools
○ Sprint times
○ Profiles and responsibilities
● Objective:
○ Manage knowledge efficiently
○ Optimize time

anyoneai.com
Fundamental Data Science
Methodology

The methodology consists of 10


steps that form an iterative
process to discover insights.

anyoneai.com
4. Exploratory Data
Analysis (EDA)

Models

Brands

anyoneai.com
(P3) = ~72 images 273 models 43 Brands
b). Problems Found
The result of the graphs shows
us that there is a positive bias of
models, so only 4% of the total
images were taken for the final
model, improving the final
accuracy.

c). Models
To create our classifier we had
to investigate CNNs , which are
networks useful for finding
patterns in images, object, face
and scene recognition

anyoneai.com
c. YOLOv7
● Real-time object detection system

● Speed and accuracy


● Clean our dataset from non-vehicles

anyoneai.com
d. Remove Background

● Improve data quality


● Better model
● Dataset reduced from 27 GB to 14 GB

anyoneai.com
5. Main Model Selected: DenseNet

● A Densenet architecture is composed of Dense blocks.

● A Dense block is a group of layers densely connected together.

● A single layer has batch normalization, ReLu activation and a 3x3 convolution.

● The feature maps inside each block are the same size

● There is a transition layer between Dense blocks.

anyoneai.com
Main Model Selected: DenseNet

anyoneai.com Source: https://arthurdouillard.com/post/densenet/


5. Main Model Selected: Model details

● Densenet201 ~ 14 million parameters, pretrained on imagenet.


● We included a dropout layer and use data augmentation.
● All layers were trained.
● 273 classes
● ~ 72 thousand images without background used for training.

anyoneai.com
5. Main Model Selected:
a Metrics

● Test accuracy: 85%


● Validation accuracy: 86%
● Train accuracy: 94%
● Training time: ~10 hours

anyoneai.com
Project Demo

anyoneai.com
Conclusions

● With a hardware Mac M1 16GB, the response time of the app was on average between 1 up
to 5 seconds.
● The model manages to distinguish between different makes and models of cars.
● The combination of models allows to distinguish the images that are not cars.
● However, using multiple models results in increased response times.
● Despite the low quality of the data, the model predicts with an accuracy greater than 85%.

anyoneai.com
Next Steps

● Increase the amount of data by adding images of vehicles.


● Expand the number of classes with vehicles with different released years and/or different
perspectives.
● Improve the quality of images with sharpening filters.
● Reduce imbalanced classes in the dataset with more samples using web crawling.
● Experiment with different hyperparameters to improve the model’s accuracy.
● Perform more data augmentation techniques.
● Try other models and architectures.
● Optimize source code to reduce response time.

anyoneai.com
Q&A

anyoneai.com

You might also like