pytorch object-detection deep-learning ssd300 vgg11 machine-learning neural-network artificial-neural-networks

SSD300-VGG11-on-Pascal-VOC-2005-Data

This project trains a SSD300 with VGG11 base on the PASCAL VOC 2005 dataset using the PyTorch deep learning framework.

Introduction

This project trains a SSD300 with VGG11 base on the PASCAL VOC 2005 dataset using the PyTorch deep learning framework.

The VOC 2005 dataset mainly contains 4 classes for detection purposes. They are:

Bicycle.
Car.
Motorbike.
Person.

This is mainly intended for beginners in deep learning who are trying to get their hands on object detection and do not want to work on very huge datasets.

Feel free to clone, download, change, distribute, and use the code any way you want.

Project Directory Structure

Follow the below given structure while cloning this repo or setting up the project and you should be good to go,

├───VOCdevkit
└───voc_2005
    ├───voc2005_1
    │   ├───Annotations
    │   ├───GTMasks
    │   └───PNGImages
    └───voc2005_2
        ├───Annotations
        ├───GTMasks
        └───PNGImages
├───logs
├───model_checkpoints
├───outputs
└───src
    │   create_data.py
    │   datasets.py
    │   detect.py
    │   detect_vid.py
    │   eval.py
    │   model.py
    │   my_utils.py
    │   train.py
    │   utils.py

If you follow the above structure, then you can just clone and run the files according to the execution commands.

PyTorch Version

PyTorch >= 1.4

Running the Code

Execute the following commands in the given order.

Note: For testing purposes, put all the image and video files inside input/test_data that you want to test or detect on.

Run the following command only once. This prepare the train and test JSON files inside the input folder.
- python create_data.py --input1 ../input/VOCdevkit/voc_2005/voc2005_1 --input2 ../input/VOCdevkit/voc_2005/voc2005_2 --save-path ../input
To train the SSD model on the dataset.
- python train.py --input ../input --continue-training yes
- If you are training for the first time, then give the --continue-training argument as no. This will train the model from scratch.
- If you have trained the model for some epochs and want to continue training, then give the --continue-training argument as yes.
- Check the number of iterations to train for inside the train.py file.
To run the model and detection on videos.
- python detect_vid.py --input ../input/test_data/video1.mp4 --resize yes
- In the above command, if --resize is yes, then the code will resize (make them smaller) the video frames only if the original frames' width is more than 800px.
To detect objects in an image.
- python detect.py --input ../input/test_data/fire_fighter.jpg
To run evaluation on validation data.
- python eval.py