SSD: Single Shot Multibox Detector, PyTorch

This is PyTorch implementation of SSD Paper.
The representative implementations based on PyTorch are amdegroot and sgrvinod, and both repositories are referenced as needed.
See Comparison of both repositories for differences between the two repositories.

Getting started

Sample results

Table of contents

• Installation
• Datasets
• Train
• Inference
• Evaluation
• Arguments
• Performance

Installation

• python>=3.6
• PyTorch>=0.4
• PIL>=6.2.1
• torchvision>=0.5.0
• matplotlib>=3.1.0

Datasets

• Download all of the PASCAL VOC 2007 and 2012 datasets.
  For consistency, the directory structure of both datasets must be matched as follows.

• Data root directory
  ├── VOC 2007
  │ └── Annotations
  │ └── ImageSets
  │ └── JPEGImages
  │ └── SegmentationClass
  │ └── SegmentationObject
  ├── VOC 2012
  │ └── Annotations
  │ └── ImageSets
  │ └── JPEGImages
  │ └── SegmentationClass
  │ └── SegmentationObject

Train

You must manually modify the root path of the VOC dataset (dataroot) in argparser.py.
Then, you can train as below.

python train.py  

Otherwise, there is a way to give it as argument as below (To differentiate between two datasets, use only ",").

python train.py --dataroot=/media/dataset/VOC2007,/media/dataset/VOC2012  

The trained model weights and loss graphs created during the training are stored in the ./results/weights/ and ./results/loss folders, respectively.

Inference

python detect_image.py --trained-model="trained_model_name.pth" --test-image="test_image_name_with_absolute_path" 

The result images are saved in the ./results/images/ folder with the p_ prefix in front of the original image name.

Evaluation

python calc_mAP.py --trained-model="trained_model_name.pth"

Arguments

• Many parameters are required to train and evaluate the object detector.
  For this, the arguments used for training and evaluation are defined separately in argparser.py.
  By using this, you can easily use and tune the parameters.
• The above-described execution examples (Train, Inference, Evaluation) are used only the minimum parameters required for operation, and all of the rest parameters are set as default.

Performance

VOC 2007 test	Achieved	Reported paper	Trained model
mAP (%)	76.5	77.2	download

Comparisons

• The weights of the pre-trained VGG16 used in the amdegroot repository and the weights of the sgrvinod directly converting the FC layer weights to the weights of the convolution layer are different.

• When sgrvinod initializing the optimizer according to the original caffe repo the learning rate for bias terms was set to twice, but in the amdegroot implementation, there is no such information.

• There is a slight difference in the size of the prior boxes made in each repository.

• There is some difference in the calculated loss by each repository (There is a difference in how the loss is calculated.).

• sgrvinod's repository is written for easy understanding and seems to follow the official paper more strictly, but the implementation of amdegroot is more optimized to PyTorch library.
  (Especially, it is prominent in time-consuming Non-maximum suppression (NMS) step)

• Therefore, I combined the two repository implementations as needed.

Feedback on anything wrong is always welcome!