tree: 2addbada44a8609782b0161b3f885e01914e5263 [path history] [tgz]
  3. config/
  4. data/
  5. dataset/
  8. detect/
  10. evaluate/
  11. model/
  12. symbol/
  13. tools/
  15. train/

SSD: Single Shot MultiBox Object Detector

SSD is an unified framework for object detection with a single network.

You can use the code to train/evaluate/test for object detection task.


This is a re-implementation of original SSD which is based on caffe. The official repository is available here. The arXiv paper is available here.

This example is intended for reproducing the nice detector while fully utilize the remarkable traits of MXNet.

  • Model converter from caffe is available now!
  • The result is almost identical to the original version. However, due to different implementation details, the results might differ slightly.

Due to the permission issue, this example is maintained in this repository separately. You can use the link regarding specific per example issues.

What's new

  • Added multiple trained models.
  • Added a much simpler way to compose network from mainstream classification networks (resnet, inception...) and Guide.
  • Update to the latest version according to caffe version, with 5% mAP increase.
  • Use C++ record iterator based on back-end multi-thread engine to achieve huge speed up on multi-gpu environments.
  • Monitor validation mAP during training.
  • More network symbols under development and test.
  • Extra operators are now in mxnet/src/operator/contrib.
  • Old models are incompatible, use e06c55d or e4f73f1 for backward compatibility. Or, you can modify the json file to update the symbols if you are familiar with it, because only names have changed while weights and bias should still be good.

Demo results

demo1 demo2 demo3


ModelTraining dataTest datamAPNote
VGG16_reduced 300x300VOC07+12 trainvalVOC07 test77.8fast
VGG16_reduced 512x512VOC07+12 trainvalVOC07 test79.9slow
Inception-v3 512x512VOC07+12 trainvalVOC07 test78.9fastest
Resnet-50 512x512VOC07+12 trainvalVOC07 test78.9fast


VGG16_reduced 300x300TITAN X(Maxwell)v5.11695
VGG16_reduced 300x300TITAN X(Maxwell)v5.1895
VGG16_reduced 300x300TITAN X(Maxwell)v5.1164
VGG16_reduced 300x300TITAN X(Maxwell)N/A836
VGG16_reduced 300x300TITAN X(Maxwell)N/A128

Forward time only, data loading and drawing excluded.

Getting started

  • You will need python modules: cv2, matplotlib and numpy. If you use mxnet-python api, you probably have already got them. You can install them via pip or package managers, such as apt-get:
sudo apt-get install python-opencv python-matplotlib python-numpy
  • Build MXNet: Follow the official instructions
# for Ubuntu/Debian
cp make/ ./
# enable cuda, cudnn if applicable

Remember to enable CUDA if you want to be able to train, since CPU training is insanely slow. Using CUDNN is optional, but highly recommended.

Try the demo

# cd /path/to/mxnet-ssd
python --gpu 0
# play with examples:
python --epoch 0 --images ./data/demo/dog.jpg --thresh 0.5
python --cpu --network resnet50 --data-shape 512
# wait for library to load for the first time
  • Check python --help for more options.

Train the model

This example only covers training on Pascal VOC dataset. Other datasets should be easily supported by adding subclass derived from class Imdb in dataset/ See example of dataset/ for details.

  • Download the converted pretrained vgg16_reduced model here, unzip .param and .json files into model/ directory by default.
  • Download the PASCAL VOC dataset, skip this step if you already have one.
cd /path/to/where_you_store_datasets/
# Extract the data.
tar -xvf VOCtrainval_11-May-2012.tar
tar -xvf VOCtrainval_06-Nov-2007.tar
tar -xvf VOCtest_06-Nov-2007.tar
  • We are going to use trainval set in VOC2007/2012 as a common strategy. The suggested directory structure is to store VOC2007 and VOC2012 directories in the same VOCdevkit folder.
  • Then link VOCdevkit folder to data/VOCdevkit by default:
ln -s /path/to/VOCdevkit /path/to/mxnet/example/ssd/data/VOCdevkit

Use hard link instead of copy could save us a bit disk space.

  • Create packed binary file for faster training:
# cd /path/to/mxnet/example/ssd
bash tools/
# or if you are using windows
python tools/ --dataset pascal --year 2007,2012 --set trainval --target ./data/train.lst
python tools/ --dataset pascal --year 2007 --set test --target ./data/val.lst --shuffle False
  • Start training:
# cd /path/to/mxnet/example/ssd
  • By default, this example will use batch-size=32 and learning_rate=0.002. You might need to change the parameters a bit if you have different configurations. Check python --help for more training options. For example, if you have 4 GPUs, use:
# note that a perfect training parameter set is yet to be discovered for multi-GPUs
python --gpus 0,1,2,3 --batch-size 32

Evalute trained model

Make sure you have val.rec as validation dataset. It's the same one as used in training. Use:

# cd /path/to/mxnet/example/ssd
python --gpus 0,1 --batch-size 128 --epoch 0

Convert model to deploy mode

This simply removes all loss layers, and attach a layer for merging results and non-maximum suppression. Useful when loading python symbol is not available.

# cd /path/to/mxnet/example/ssd
python --num-class 20

Convert caffe model

Converter from caffe is available at /path/to/mxnet/example/ssd/tools/caffe_converter

This is specifically modified to handle custom layer in caffe-ssd. Usage:

cd /path/to/mxnet/example/ssd/tools/caffe_converter
python deploy.prototxt name_of_pretrained_caffe_model.caffemodel ssd_converted
# you will use this model in deploy mode without loading from python symbol(layer names inconsistent)
python --prefix ssd_converted --epoch 1 --deploy

There is no guarantee that conversion will always work, but at least it's good for now.

Legacy models

Since the new interface for composing network is introduced, the old models have inconsistent names for weights. You can still load the previous model by rename the symbol to and call with python train/ --network legacy_xxx For example:

python --network '' --prefix model/ssd_300 --epoch 0