A Pytorch Implementation of RetinaNet for object detection

Introduction

This project is a pytorch implementation of RetinaNet. During the implementing, I referred several implementations to make this project work:

• kuangliu/pytorch-retinanet, this repository give several main scripts to train RetinaNet, but doesn't give the results of training.
• fizyr/keras-retinanet, this repository completely give the training, test, evaluate processes, but it is based on Keras.
• roytseng-tw/Detectron.pytorch, this repository is a implementation of Detectron based on pytorch, but it doesn't support RetinaNet in this moment.

For this implementation, it has the following features:

• It supports multi-image batch training. Change the original sampler into MinibatchSampler to support different multi-image size in minibatch.
• It supports multiple GPUs training. Change the original DataParallel in Pytorch to support minibatch supported dataset.

Results of RetinaNet

Now, I get the result using COCOAPI, the training AP is 33.1, compare to 34.0 in the original paper, the result is comparable. And the details of AP are as follow:

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.331
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.500
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.353
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.154
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.371
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.459
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.279
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.429
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.459
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.238
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.520
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.598

Getting Started

Clone the repo:

git clone https://github.com/wsnedy/pytorch-retinanet.git

Requirements

Tested under python2.7.

• python packages
  • pytorch=0.3.1
  • torchvision=0.2.0
  • matplotlib
  • numpy
  • opencv
  • pycocotools — for COCO dataset, also available from pip.
• An NVIDAI GPU and CUDA 8.0 or higher. Some operations only have gpu implementation.

Get pretrained model

Download pretrained ResNet50 params from the following url.

mkdir pretrained_model
cd pretrained_model
wget https://download.pytorch.org/models/resnet50-19c8e357.pth
mv resnet50-19c8e357.pth resnet50.pth

Get the pretrained RetinaNet by run the script:

cd network
python get_state_dict.py

Data Preparing

Download the coco images and annotations from coco website.

And make sure to put the files as the following structure:

coco
├── annotations
|   ├── instances_minival2014.json
│   ├── instances_train2014.json
│   ├── instances_val2014.json
│   ├── instances_valminusminival2014.json
│   ├── ...
|
└── images
    ├── train2014
    ├── val2014
    ├── ...

When training, change the root path to your own data path.

Training model

For the hyper-parameters, I just put them in the scripts, and I will put all the hyper-parameters in a config file. The setting is as follows:

• For multi-gpus, it will use all the available gpus in default. change the device_ids in bin/train.py if you want to specific gpus.
• For batch_size, I use batch_size = 24, if you want to change, you have to change two places, batch_size=24 and iteration_per_epoch = int(len(dataloader) / 24.) in bin/train.py
• For img_per_minibatch, I use img_per_minibatch = 3 to achieve batch_size=24, change it if you want to use other minibatch number.
• To do, put all the hyper-parameters into config file.

Training from scratch

Training RetinaNet using following code, and after each epoch, it will give a evaluation in minival2014 dataset:

python bin/train.py

Training from a pretrained checkpoint

If you want to load the checkpoint, use the follow code. If you have more than one checkpoint, change the code checkpoint = torch.load('../checkpoint/ckpt.pth') in train.py to load different checkpoint:

python bin/train.py --resume

Inference

• To do: demo.py

Support Dataset

Only COCO supported now, for different dataset, change a little bit in datasets will be work.

• To do: support VOC dataset.