This is my first object detection project built and run locally using SSD, with references to the implementation of SSD from amdegroot and huutrinh68 with minor adjustment. SSD is an old object detection algorithm, released in 2016, but it still covers a lot of useful ideas of how an object detection algorithms may work.
- Data Processing
- SSD Modules
- Usage
Here, I am using the VOC2012 dataset for training my SSD model. You can download it automatically by using the prepare_data.py, which will create a data folder and install the VOC2012 dataset from the internet and untar the dataset (in case the dataset does not exist).
python prepare_data.py
After having our dataset, we continue to use make_data_path.py (which will create lists, each containing images and corresponding annotation links for the later process). The extract_anno_info.py extracts information from annotation links, which are in xml format. The transform.py uses some basic transformations for data augmentation. All of the 3 files will then be used in the dataset.py, which will create train and val DataLoaders containing images after transformation and the annotations (bounding boxes coordinate, class) for training the model.
The SSD Modules contain VGG layers, the extra feature layers, the L2Norm, and loc and conf layers.
View the model.py for more architecture information
Firstly, clone this repository by executing:
git clone https://github.com/TaiQuach123/Object-Detection-with-SSD.git
Then, get the pre-trained weights for the VGG part as initial weights to help the model converge faster
cd './data'
mkdir 'weights'
cd './weights'
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
cd ../..
Finally, executing the train.py:
python train.py