https://github.com/coldestadam/fruit-fasterr-cnn-detection

Using Pytorch to create multiple Faster R-CNN models to detect images of fruit.
https://github.com/coldestadam/fruit-fasterr-cnn-detection

Last synced: 1 day ago
JSON representation

Using Pytorch to create multiple Faster R-CNN models to detect images of fruit.

• Host: GitHub
• URL: https://github.com/coldestadam/fruit-fasterr-cnn-detection
• Owner: Coldestadam
• Created: 2020-08-02T05:27:33.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2022-10-25T01:58:38.000Z (over 2 years ago)
• Last Synced: 2024-12-03T16:33:52.888Z (about 2 months ago)
• Language: HTML
• Size: 8.35 MB
• Stars: 1
• Watchers: 1
• Forks: 2
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Fruit Detection using Faster R-CNNs

Using Pytorch to create multiple Faster R-CNN models to detect images of fruit.

## Dataset

[Fruit Images for Object Detection](https://www.kaggle.com/mbkinaci/fruit-images-for-object-detection)

This dataset has images of apples, bananas, and oranges for the purpose of object detection. There are also images that have all three fruits within them, but most are uniform in their classes.

## Modified Faster R-CNN model

I modified PyTorch's torchvision Faster R-CNN class to both give losses and predictions while in eval() mode to allow ease of access to calculate the training loss.

My modified torchvision repository is [HERE](https://github.com/Coldestadam/vision_)

These are the files I modified:

1. [generalized_rcnn.py](https://github.com/Coldestadam/vision_/blob/master/torchvision/models/detection/generalized_rcnn.py)

2. [roi_heads.py](https://github.com/Coldestadam/vision_/blob/master/torchvision/models/detection/roi_heads.py)

3. [rpn.py](https://github.com/Coldestadam/vision_/blob/master/torchvision/models/detection/rpn.py)

## Methods

I chose to learn more about other optimization algorithms and I followed the advice of this article named [*Optimization Algorithms in Deep Learning*](https://towardsdatascience.com/optimization-algorithms-in-deep-learning-191bfc2737a4). The writer states that the [Adam Optimizer](https://en.wikipedia.org/wiki/Stochastic_gradient_descent#Adam) is the best algorithm for its ability to both use momentum and scaling to mix the algorithms of SGD + momentum and RMS Prop.

I used the writers advice to define the hyperparameters:

- B1 = 0.9

- B2 = 0.999

- Learning Rate = 0.001–0.0001

## Samples

Here is a sample of my Faster R-CNN model predicting the correct location and class of the orange.

![img](samples/sample_9.jpg)

[Please click here to see my 10 sample images.](samples)

## Results:

I was not able to get the best results due to lack of resources, but I was able to implement the algorithm which was my goal. Thank you for checking this out!