SSD: Single Shot MultiBox Detector

This is an implementation of SSD (Single Shot MultiBox Detector) using Chainer

Performance

Pascal VOC2007 Test

Method	Original	Test only (model conversion)	Train and Test
SSD300	77.6 %	77.5 %	77.3 %
SSD512	79.5 % [3]	79.6 %	80.4% *

Requirements

• Python 3.5+
• Chainer 1.24+
  • ExponentialShift had a resuming issue in older version.
• OpenCV 3
• Matplotlib

Usage

Testing

1. Download pre-traind Caffe model from https://github.com/weiliu89/caffe/tree/ssd#models

$ curl -LO http://www.cs.unc.edu/%7Ewliu/projects/SSD/models_VGGNet_VOC0712_SSD_300x300.tar.gz
$ tar xf models_VGGNet_VOC0712_SSD_300x300.tar.gz

2. Convert weights

$ ./caffe2npz.py models/VGGNet/VOC0712/SSD_300x300/VGG_VOC0712_SSD_300x300_iter_120000.caffemodel VGG_VOC0712_SSD_300.npz

3.a Test and Eval with Pascal VOC dataset

$ curl -LO http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
$ tar xf VOCtest_06-Nov-2007.tar
$ ./test.py [--arch 300] VGG_VOC0712_SSD_300.npz 2007-test [--gpu gpu]
(result/comp4_det_test_*.txt will be generated)
$ ./eval.py 2007-test
(calculate unofficial mAP)

3.b Test with an image

$ ./demo.py [--arch 300] VGG_VOC0712_SSD_300.npz VOCdevkit/VOC2007/JPEGImages/000001.jpg
5 0.0130746 273 86 293 167
5 0.0113751 140 208 195 261
9 0.0211564 82 444 117 484
9 0.0200858 3 27 343 492
9 0.0164014 68 446 97 486
9 0.0158782 59 425 131 488
9 0.0157032 128 455 187 488
9 0.0131429 8 335 63 427
12 0.82896 51 240 202 374
12 0.147764 8 227 237 476
12 0.0550958 11 1 350 498
15 0.985803 12 5 354 492
15 0.0196945 273 95 294 173
15 0.0134011 274 92 319 184
18 0.0143462 11 1 350 498

Training

1. Download pre-trained VGG16 model (fc reduced) from https://gist.github.com/weiliu89/2ed6e13bfd5b57cf81d6

$ curl -LO http://cs.unc.edu/~wliu/projects/ParseNet/VGG_ILSVRC_16_layers_fc_reduced.caffemodel

2. Convert weights

$ ./caffe2npz.py VGG_ILSVRC_16_layers_fc_reduced.caffemodel VGG_ILSVRC_16_fc_reduced.npz

3. Download VOC dataset

$ curl -LO http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
$ curl -LO http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
$ tar xf VOCtrainval_06-Nov-2007.tar
$ tar xf VOCtrainval_11-May-2012.tar

4. Train

$ ./train.py --init VGG_ILSVRC_16_fc_reduced.npz --train 2007-trainval --train 2012-trainval [--gpu gpu]

SSD512 train environment

NVIDIA GeForce® GTX 1080 Ti, Batch size: 8, 480k iterations

ToDo

• Multi GPUs support

References

• [1]: Wei Liu, et al. "SSD: Single shot multibox detector" ECCV 2016.
• [2]: Original implementation
• [3]: Cheng-Yang Fu, et al. "DSSD : Deconvolutional Single Shot Detector" arXiv 2017.