This is a fast and concise implementation of Faster R-CNN with TensorFlow2 based on endernewton TensorFlow1 implementation and other works.

The purpose and features of this repository:

• Recurrence of origin paper <Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks> with superior performance.
• Concise ,flat and straightforward model architecture with abundant comments.
• Extend from tf1 to tf2 with Eager implementation and Graph Execution.
• Faster CUDA Non Maximum Suppression implementation with Cython wrapper.
• A good start point for further inspection of classic two-stage object detection architecture.

In order to be compatible with original setup, this result shown here is initialized with VGG16 ImageNet pretrained model and trained on Pascal VOC2007 trainval and tested on test dataset. The inference may need warm up to achieve the best speed.

1. Clone this repository

   git clone 
   cd tf2-faster-rcnn

2. Install denpendencies with virtual env anaconda This project is trained and tested under Python3.7, TensorFlow v2.2 and cudatoolkit 10.1.

   conda create -n tf2 tensorflow-gpu cudatoolkit=10.1 
   conda activate tf2
   conda install numpy opencv cython scipy lxml

3. Build CUDA NMS with Cython wrapper. This implementation uses CUDA NMS as default. i. Check your GPU's compute capability and modify -arch=sm_86 in the build.bat. Here uses sm_86 as an example (-Xptxas="-v" is optional for showing register and smem usage):

   nvcc -lib -O3 -arch=sm_86 -Xptxas="-v" -o nms_kernel.lib gpu_nms.cu

   ii.Run the following commands:

   cd model/utils/nms/gpu
   ./build.bat

   Compared with pure Python NMS and Cython implementation, the gpu version is ~15x and ~12x faster in the setting of 6000 bboxes as input. Differ from the original(1-D), this CUDA NMS kernel is implemented by expanding IoU calculation into 2-D and parallelized in each CUDA thread.

4. Build the Cython CPU NMS (optional)

   cd model/utils/nms/cpu
   python build.py build_ext --inplace

1. Prepare Pascal VOC2007 dataset Download VOCdevkit, VOCtrainval, VOCtest and extract all into data/VOCdevkit.

   - VOCdevkit
        - VOC2007
        - VOCcode
        - ...
   

2. Download VGG16 ImageNet pretrained model(optional) If you want to initialize the training with ImageNet pretrained model and train from scratch, download vgg16 to model/vgg16.h5.

3. Download Pascal VOC2007 trainval dataset trained model weights to model/ckpt/ from Google Drive.

4. Start training

   python train.py 

   Optional arguments:

   • --scratch: Use ImageNet pretrained vgg16 to train the model from scratch.
   • --epoch <n>: n = the number of epochs. Each epoch will train over trainval dataset once.
   • --record_all: Include kernel and gradient information in tensorboard summary.

   Training results viewed by tensorboard while training:

   tensorboard --logdir model/logs --bind_all

   

If you start the training first time, it will take some time to set up the dataset(preprocessing). After preprocessing the directory will contain data/cache/voc_07_train_gt_roidb.pkl, data/cache/voc_07_test_gt_roidb.pkl.

• Test with VOC2007 test dataset:

  python test.py -i voc07_test

• Test with self defined images: Put your images under demo/images

  python test.py -i demo

• Optional arguments
  • --unpop: run test without popping up result images as a new window
  • --save: save result images to demo/results

Examples:

image source: Internet

This project is built based on the following great works:

• endernewton's tf1 implementation
• py-faster-rcnn by Ross Girshick
• chenyuntc's pytorch implementation

Please check out the model structure below before digging into details. ;)

Also feel free to pull requests and open an issue if you have any problem!

Faster R-CNN Model Structure: