SSD Keras - Object detection

Thanks rykov8 for basic structure and wikke for updating model. I split from generate_data to model into several part which notebook name starts with TX_ . My goal is sharing my knowlege in simple way to everyone. Have fun !

Requirement

• tensorflow >= 1.4.0
• opencv-python
• keras >= 2.0
• scipy
• matplotlib
• google.protobuf

[Notice for Dataset]

you must to create dict classes, check sample.

Dataset

• VOC2007 : wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
• VOC2012 : wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar

Pretrained Model Weights

download link is here

Chief Concept

SSD as its name aims to detect bounding-box in single-shot, nontheless how we get these bounding-box and its category, here is solution.

1. Bounding-Box Detection: We use CNN as our model for training, but question is how we train it?
   • Q1: How do we train it with CNN model?
   • A1: Everytime we want to train a model we need to supply ground-truth data, which in these model we apply PriorBox to encoding ground-truth bbox to feature map data, and then its get several datas pertaining to feature scaling ratio(aspect ratio) and image scaling ratio(resolution ratio). Furthermore, we can use aforementioned data as training data to train our model.(see notebook: T3_AssignBBoxes.ipynb, T4_PriorBox.ipynb, and T7_LossFunction) if you want to know how the formation of priorbox which is used in T3_AssignBBoxes.ipynb, you can check T6_CreatePriorBoxes.
2. Classification:
   • Q1:
   • A1:
3. How to comebine above into one.

Description

1. In T1_GenerateData.ipynb shows you each of steps image process effect and flow of generating dataset.
2. In T2_Modulized_GenerateData.ipynb shows you I modulized generate_data then placed in utils/generate_data.py and demo part.
3. In T3_AssignBBoxes.ipynb describes the way to assign bboxes which is core procedure in generate_data.py.
4. In T4_PriorBox.ipynb describe the layer in SSD model which is core layer for Object-detection.
5. In T5_SSDModel.ipynb I restructure original model which I think more easily to understand.
6. In T6_CreatePriorBoxes I find this on website and make a explanation.
7. In T7_LossFunction Note here, I still have to figure it out.
8. In T8_DecodePredictValue As the filename, script is quite intuition, very easy to understand.