chenzx921020 / mxnet-ssd

based on mxnet， ssd

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 your own object detection task.

Disclaimer

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.

Compared with official version in mxnet, this repo provides the process of sample generating,training and forward predicting and continuously updating

• Basic preparation: Before bellow steps, you need to build the directory model and save the pretrained model in it

# mkdir model
# cd model
hadoop fs -get hdfs://hobot-bigdata/user/chenzhixuan/ssd_pretrained_model/vgg16-reduced-*

• Sample generating: In dataset/pascal_voc.py, we have modified the data import ways. Firstly, read the txt data and output lst file, and secondly, compress lst file into rec file with mxnet tools im2rec.py
  The txt data format: img_url lefttop_x lefttop_y rightbottom_x rightbottom_y

sh prepare_dataset.sh

The --root is the source txt, the --target is the target lst, you can modify them by your convenience.
In im2rec process, you need provide the source image path and lst file above.

• Start training:

python train.py

• By default, this example will use batch-size=32 and learning_rate=0.004. You might need to change the parameters a bit if you have different configurations.
  Check python train.py --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-gpu
python train.py --gpus 0,1,2,3 --batch-size 128 --lr 0.001

• Attention memory: training ssd300 in vgg16-reduced will comsume at least 2 gpus, and ssd500 comsuming 24g is not provided.

Convert model to deploy mode

This simply removes all loss layers, and attach a layer for merging results and non-maximum suppression.
Rename the model and symbol to ssd_vgg16_reduced_300-symbol.json and ssd_vgg16_reduced_300-000x.params. The x is up to epochs.

python deploy.py --prefix ssd_vgg_reduced_300 --epoch x --num-class 2 

After steps above, in path ./model will generate the two files with prefix deploy

python demo.py --image 3304.png --prefix model/deploy_ssd_vgg16_reduced_300 --epoch x --deploy --data-shape 300

The data-shape must be consistent with training input size