namoshika/ssd_keras

Port of Single Shot MultiBox Detector to Keras

SSD: Single Shot MultiBox Detector on TensorFlow 2.0

Implemented SSD300 in TensorFlow 2.0 (keras API).
This SSD300 has some improvements over the forked repository.

1. Port to TF2.0 (Eager Execution + Keras API)
2. Evaluate ROC curves, mAP Score.
3. Remove predatas (gt_pascal.pkl, prior_boxes_ssd300.pkl).
4. Remove output feature which didn't have trainable weights.

Requirements

• Python v3.7
• TensorFlow v2.0
• python_voc_parser v1.0.0
• imgaug v0.3.0

Usage

1. Open demo with VSCode or Jupyter.
   • sample_demo.py: You can try predict and training. My work space (VSCode friendly. I'm attached to breakpoint. XD).
   • sample_demo.ipynb: This is same as sample_demo.py.
2. Run # データセットをダウンロード cell.
   Some minutes after... Downloaded the PascalVOC 2007 onto ./data.

   ./data/VOCdevkit/VOC2007/
                        |- Annotations
                        |- ImageSets
                        |- JPEGImages
                        |- SegmentationClass
                        |- SegmentationObject
   

3. Download the weights_SSD300.hdf5 onto ./data. This is weights was ported from the original models.

   ./data/weights_SSD300.hdf5
   

4. Run following cells... XD

Copyright

Copyright (c) 2020 namoshika

This repository is forked https://github.com/rykov8/ssd_keras.
Copyright (c) 2016 Andrey Rykov