Faster RCNN

Faster R-CNN Resnet-101-FPN implementation based on TensorFlow 2.0.

Requirements

• python 3.*
• tensorflow>=2.2.0rc3
• scikit-image
• cv2

Training

args

-b batch_size (default = 2)

-f finetune (default = 0)

-l learning_rate (default = 1e-4)

-e epochs (default = 100)

-c checkpoint (default = 1 in train_epoch, default = 500 in train_batch)

-n normalization (default = ImageNet's mean and std.)
   n = 0: no normalization;
   n = 1: Company Articles Dataset's mean and std.

Command

python train_batch.py [commands]

Wiki

• Project Design Documentation
• Records on Debug & Addition of Original Project
• Visualized comparison between different normalization scheme
• Detailed Training and Testing Logs of Training Without Normalization
• Analysis on training result caused by different normalization scheme I
• Analysis on training result caused by different normalization scheme II ☆
• Analysis on training result caused by different anchor ratios
• Comparison with Detectron's Faster RCNN

Experiment

Comparison between without normalization and ImageNet Normalization

Comparison between Anchor Ratio of (0.5 1 2), (1 2 4) and (1 3 6)

Comparison between TensorFlow2 FasterRCNN and PyTorch Detectron2 FasterRCNN

Acknowledgement

This work builds on many excellent works, which include:

• Heavily based on tf-eager-fasterrcnn
• matterport/Mask_RCNN
• open-mmlab/mmdetection