Faster RCNN on TF2

This repo modify matterport/Mask_RCNN for implementation of Faster RCNN model

Architecture

setup

• Anaconda on Windows 10
• Python 3.6.6
• Use Spyder IDE edit my code
• GTX 1060 (Laptop)
• CUDA 10.1 + cudnn v8.0.4.30
• RAM: 16GB
  The code can run on tf2.3

Work list

• Remove mask part
• Input Shapes dataset
• Input BCCD dataset
• Input VOC 2007 data
• Input Custom data
• Input multi voc format dataset
• Input COCO data
• Change backbone (use Keras application)
• Head layer reduce weights
• Inspect training results (model output and weights)
• Training results
• Data augmentation

Tips

• Install package can follow here
• Input data

  • Modify config
    IMAGE_RESIZE_MODE = "none" or "square"
    "square" can handle different size image

    NUM_CLASSES
    Set number of classes of custom dataset

  • Use VOC annotation format (.xml)

Results

Shapes dataset
mAP: 0.95 (100 samples)

VOC dataset
mAP: 0.49 (4952 samples)

BCCD dataset
mAP: 0.74 (72 samples)

Custom dataset (Digital Holographic Microscopy)

Refer

Paper: here
PSACAL VOC 2007 dataset: here
BCCD dataset: here
Hand dataset on my repo
Related work on my repo