tanakataiki / ssd_kerasV2

BetterNetWork4SSD

Keras Implemention of CustomNetwork-SSD

This Work is going to support

MobileNet-SSD

VGG16-SSD300

VGG16-SSD512

FeatureFused-SSD300

Xception-SSDLite (Tanaka Original Ver)

MobileNetV2-SSDLite

Video is from free to use https://www.pexels.com/video/a-day-in-the-park-1466210/

I set threshold 0.9 to ignore wrong detection but usually thresh=0.6 So Please do not try this at home(this doesnt affect loss or map at all)

Requirements

This code was tested with Keras v2.1.5, Tensorflow v1.6.0 GTX1080 Tensorflow・Keras・Numpy・Scipy・opencv-python・pillow・matplotlib・h5py

My Weights Are Available From Here and WELCOME to upload your fine tuned weights

https://drive.google.com/drive/u/0/folders/1F8GjD3BFhf_hv9Ipez0twRptYc3P8YwP

Please write loss, acc and if possible mAp and your name if you want as your weight name https://drive.google.com/drive/folders/1u-INV0pNjSjwNgbupXVpr1lwEsTMKW3F?usp=sharing

Pull Request Is always welcome

As the truely perfect model doesn't exist forever there is still a way better. (currently I don't have enought time to search very deep into details too...)

To Whom

I use this for a detection of few categories and simple shape detection (for my purpose) but weak for coco or voc. This repository is just for my study of network architecture. So there is no measurement but for gif so dont trust too much. If you would like to use better and prooved one ,please use official version of ssd or yolo. If you want to study architecture itself without training technic ,this repo will be good to be as one of many references.

Reference

SSD : https://github.com/rykov8/ssd_keras/blob/master/ssd.py

Caffe : https://github.com/weiliu89/caffe/tree/ssd

SSD : https://arxiv.org/abs/1512.02325

FSSD : https://arxiv.org/abs/1712.00960

FFSSD : https://arxiv.org/abs/1712.00960

DSSD : https://arxiv.org/abs/1701.06659

VGG : https://arxiv.org/abs/1409.1556

MobileNet : https://arxiv.org/abs/1704.04861

MobileNetV2 : https://arxiv.org/abs/1801.04381

Xception : https://arxiv.org/abs/1610.02357

MobileNetSSD : https://github.com/chuanqi305/MobileNet-SSD

MobileNetV2-SSDLite : https://github.com/chuanqi305/MobileNetv2-SSDLite

VGG16-SSD : https://qiita.com/tanakataiki/items/226c2460738361d2c4eb

MobileNet-SSD : https://qiita.com/tanakataiki/items/41509e1b0f4a9dcd01b1

FeatureFused-SSD : https://qiita.com/tanakataiki/items/36e71e7d2f5705bd98bb

Xception-SSDLite : https://qiita.com/tanakataiki/items/63fa46f529174d8e4c03

Licence

The MIT License (MIT)

Copyright (c) 2018 Taiki Tanaka