Author:

Yu Zhong yuz871@eng.ucsd.edu
Shiming Luo shl666@ucsd.edu

Requirement:

Python 3.6
Tensorflow

download (original SSD300 network check point)[https://drive.google.com/a/eng.ucsd.edu/file/d/1znGTqsAa0M9gN1EDCqP9Pw667kjAsgwj/view?usp=sharing]
download (SSD300 added block3 check point)[https://drive.google.com/a/eng.ucsd.edu/file/d/12NkA9rBi_4_x32Vbcbftc1SUxXPcO4Ut/view?usp=sharing]

Introduction:

This is a SSD300 based small object detection model. We add the 3rd block of the orignal SSD model and apply the method called Cropping Prediction.

The architecture of modified SSD300 model:

The flow chart of Cropping Prediction.

SSD demo

The SSD Notebook contains a demo of our work.

Here are two examples of successful detection outputs:

To run the notebook you first have to unzip the checkpoint files in ./checkpoint

unzip original_model.zip
unzip add_block3.zip

and then start a jupyter notebook with

jupyter notebook notebooks/ssd_notebook.ipynb

Datasets

VOC 2007
VOC 2012
WIDER FACE
The way to convert VOC 2007 and VOC 2012 plz see balancap/SSD-Tensorflow

The way to convert WIDER FACE plz see yeephycho/widerface-to-tfrecord

Training

The script train_ssd_network.py is in charged of training the network.

For instance, one can fine a model starting from the former as following:

DATASET_DIR=./tfrecords/WF/train
TRAIN_DIR=./log/WF_original
python train_ssd_network.py \
    --train_dir=${TRAIN_DIR} \
    --dataset_dir=${DATASET_DIR} \
    --dataset_name=WiderFace \
    --dataset_split_name=train \
    --model_name=ssd_300_vgg \
    --num_classes=2 \
    --save_summaries_secs=60 \
    --save_interval_secs=600 \
    --weight_decay=0.0005 \
    --optimizer=adam \
    --learning_rate=0.001 \
    --learning_rate_decay_factor=0.94 \
    --batch_size=32

A number of pre-trained weights of popular deep architectures can be found on TF-Slim models page.

Credit

Inspired by:
-kuangliu/pytorch-ssd
-yeephycho/widerface-to-tfrecord
-balancap/SSD-Tensorflow