High quality, fast, modular reference implementation of SSD in PyTorch 1.0

This repository implements SSD (Single Shot MultiBox Detector). The implementation is heavily influenced by the projects ssd.pytorch, pytorch-ssd and maskrcnn-benchmark. This repository aims to be the code base for researches based on SSD.

Highlights

• PyTorch 1.0
• GPU/CPU NMS
• Multi-GPU training and inference
• Modular
• Visualization(Support Tensorboard)
• CPU support for inference

Installation

Requirements

1. Python3
2. PyTorch 1.0
3. yacs
4. GCC >= 4.9
5. OpenCV

Build

# build nms
cd ext
python build.py build_ext develop

Train

Setting Up Datasets

Pascal VOC

For Pascal VOC dataset, make the folder structure like this:

VOC_ROOT
|__ VOC2007
    |_ JPEGImages
    |_ Annotations
    |_ ImageSets
    |_ SegmentationClass
|__ VOC2012
    |_ JPEGImages
    |_ Annotations
    |_ ImageSets
    |_ SegmentationClass
|__ ...

Where VOC_ROOT default is datasets folder in current project, you can create symlinks to datasets or export VOC_ROOT="/path/to/voc_root".

COCO

For COCO dataset, make the folder structure like this:

COCO_ROOT
|__ annotations
    |_ instances_valminusminival2014.json
    |_ instances_minival2014.json
    |_ instances_train2014.json
    |_ instances_val2014.json
    |_ ...
|__ train2014
    |_ <im-1-name>.jpg
    |_ ...
    |_ <im-N-name>.jpg
|__ val2014
    |_ <im-1-name>.jpg
    |_ ...
    |_ <im-N-name>.jpg
|__ ...

Where COCO_ROOT default is datasets folder in current project, you can create symlinks to datasets or export COCO_ROOT="/path/to/coco_root".

Single GPU training

# for example, train SSD300:
python train_ssd.py --config-file configs/ssd300_voc0712.yaml --vgg vgg16_reducedfc.pth

Multi-GPU training

# for example, train SSD300 with 4 GPUs:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS train_ssd.py --config-file configs/ssd300_voc0712.yaml --vgg vgg16_reducedfc.pth

The configuration files that I provide assume that we are running on single GPU. When changing number of GPUs, hyper-parameter (lr, max_iter, ...) will also changed according to this paper: Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. The pre-trained vgg weights can be downloaded here: https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth.

Demo

Predicting image in a folder is simple:

python demo.py --config-file configs/ssd300_voc0712.yaml --weights path/to/trained/weights.pth --images_dir demo

Then the predicted images with boxes, scores and label names will saved to demo/result folder.

Currently, I provide weights trained as follows:

Performance

Origin Paper:

Our Implementation:

Details:

mAP: 0.7783
aeroplane       : 0.8252
bicycle         : 0.8445
bird            : 0.7597
boat            : 0.7102
bottle          : 0.5275
bus             : 0.8643
car             : 0.8660
cat             : 0.8741
chair           : 0.6179
cow             : 0.8279
diningtable     : 0.7862
dog             : 0.8519
horse           : 0.8630
motorbike       : 0.8515
person          : 0.8024
pottedplant     : 0.5079
sheep           : 0.7685
sofa            : 0.7926
train           : 0.8704
tvmonitor       : 0.7554

mAP: 0.8025
aeroplane       : 0.8582
bicycle         : 0.8710
bird            : 0.8192
boat            : 0.7410
bottle          : 0.5894
bus             : 0.8755
car             : 0.8856
cat             : 0.8926
chair           : 0.6589
cow             : 0.8634
diningtable     : 0.7676
dog             : 0.8707
horse           : 0.8806
motorbike       : 0.8512
person          : 0.8316
pottedplant     : 0.5238
sheep           : 0.8191
sofa            : 0.7915
train           : 0.8735
tvmonitor       : 0.7866