SSD: Single Shot MultiBox Object Detector, in PyTorch

A PyTorch implementation of Single Shot MultiBox Detector from the 2016 paper by Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang, and Alexander C. Berg. The official and original Caffe code can be found here.

• Note: While I would love it if this were my full-time job, this is currently only a hobby of mine so I cannot guarantee that I will be able to dedicate all my time to updating this repo. That being said, thanks to everyone for your help and feedback it is really appreciated, and I will try to address everything as soon as I can.

Table of Contents

• Installation
• Datasets
• Train
• Evaluate
• Performance
• Demos
• Future Work
• Reference

       

Installation

• Install PyTorch by selecting your environment on the website and running the appropriate command.
• Clone this repository.
  • Note: We currently only support Python 3+.
• Then download the dataset by following the instructions below.
• We now support Visdom for real-time loss visualization during training!
  • To use Visdom in the browser:

  # First install Python server and client 
  pip install visdom
  # Start the server (probably in a screen or tmux)
  python -m visdom.server

  • Then (during training) navigate to http://localhost:8097/ (see the Train section below for training details).
• Note: For training, we currently only support VOC, but are adding COCO and hopefully ImageNet soon.
• UPDATE: We have switched from PIL Image support to cv2. The plan is to create a branch that uses PIL as well.

Datasets

To make things easy, we provide a simple VOC dataset loader that inherits torch.utils.data.Dataset making it fully compatible with the torchvision.datasets API.

VOC Dataset

Download VOC2007 trainval & test

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>

Download VOC2012 trainval

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>

Training SSD

• First download the fc-reduced VGG-16 PyTorch base network weights at: https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
• By default, we assume you have downloaded the file in the ssd.pytorch/weights dir:

mkdir weights
cd weights
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth

• To train SSD using the train script simply specify the parameters listed in train.py as a flag or manually change them.

python train.py

• Note:
  • For training, an NVIDIA GPU is strongly recommended for speed.
  • Currently we only support training on v2 (the newest version).
  • For instructions on Visdom usage/installation, see the Installation section.
  • You can pick-up training from a checkpoint by specifying the path as one of the training parameters (again, see train.py for options)

Evaluation

To evaluate a trained network:

python eval.py

You can specify the parameters listed in the eval.py file by flagging them or manually changing them.

Performance

VOC2007 Test

mAP

Original	Converted weiliu89 weights	From scratch w/o data aug	From scratch w/ data aug
77.2 %	77.26 %	58.12%	77.43 %

Evaluation report for the current version

VOC07 metric? Yes

AP for aeroplane = 0.8172
AP for bicycle = 0.8544
AP for bird = 0.7571
AP for boat = 0.6958
AP for bottle = 0.4990
AP for bus = 0.8488
AP for car = 0.8577
AP for cat = 0.8737
AP for chair = 0.6147
AP for cow = 0.8233
AP for diningtable = 0.7917
AP for dog = 0.8559
AP for horse = 0.8709
AP for motorbike = 0.8474
AP for person = 0.7889
AP for pottedplant = 0.4996
AP for sheep = 0.7742
AP for sofa = 0.7913
AP for train = 0.8616
AP for tvmonitor = 0.7631
Mean AP = 0.7743

FPS

GTX 1060: ~45.45 FPS

Demos

Use a pre-trained SSD network for detection

Download a pre-trained network

• We are trying to provide PyTorch state_dicts (dict of weight tensors) of the latest SSD model definitions trained on different datasets.
• Currently, we provide the following PyTorch models:
  • SSD300 v2 trained on VOC0712 (newest PyTorch version)
    • https://s3.amazonaws.com/amdegroot-models/ssd300_mAP_77.43_v2.pth
  • SSD300 v2 trained on VOC0712 (original Caffe version)
    • https://s3.amazonaws.com/amdegroot-models/ssd_300_VOC0712.pth
  • SSD300 v1 (original/old pool6 version) trained on VOC07
    • https://s3.amazonaws.com/amdegroot-models/ssd_300_voc07.tar.gz
• Our goal is to reproduce this table from the original paper

Try the demo notebook

• Make sure you have jupyter notebook installed.
• Two alternatives for installing jupyter notebook:

  1. If you installed PyTorch with conda (recommended), then you should already have it. (Just navigate to the ssd.pytorch cloned repo and run): jupyter notebook

  2. If using pip:

# make sure pip is upgraded
pip3 install --upgrade pip
# install jupyter notebook
pip install jupyter
# Run this inside ssd.pytorch
jupyter notebook

• Now navigate to demo/demo.ipynb at http://localhost:8888 (by default) and have at it!

Try the webcam demo

• Works on CPU (may have to tweak cv2.waitkey for optimal fps) or on an NVIDIA GPU
• This demo currently requires opencv2+ w/ python bindings and an onboard webcam
  • You can change the default webcam in demo/live.py
• Install the imutils package to leverage multi-threading on CPU:
  • pip install imutils
• Running python -m demo.live opens the webcam and begins detecting!

TODO

We have accumulated the following to-do list, which you can expect to be done in the very near future

• Still to come:
  • Train SSD300 with batch norm
  • Add support for SSD512 training and testing
  • Add support for COCO dataset
  • Create a functional model definition for Sergey Zagoruyko's functional-zoo

References

• Wei Liu, et al. "SSD: Single Shot MultiBox Detector." ECCV2016.
• Original Implementation (CAFFE)
• A huge thank you to Alex Koltun and his team at Webyclip for their help in finishing the data augmentation portion.
• A list of other great SSD ports that were sources of inspiration (especially the Chainer repo):
  • Chainer, Keras, MXNet, Tensorflow