PsRoIAlign Operation In Tensorflow C++ API

PsRoIAlign involves interpolation techniques for PsRoiPooling (position-sensitive RoI pooling operation), the interpolation idea is proposed in RoIAlign to avoid any quantization of the RoI boundaries. The first adoption of PsRoIAlign might be in this paper Light-Head R-CNN: In Defense of Two-Stage Object Detector.

You can find more details about the RoiPooling technique in Fast R-CNN and Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition.

This repository contains code of the implement of PsRoIAlign operation in Tensorflow C++ API. You can use this operation in many popular two-stage object detector. Both research work using PsRoIAlign and contribution to this repository are welcomed.

For using this op in your own machine, just following these steps:

• copy the header file "cuda_config.h" from "your_python_path/site-packages/external/local_config_cuda/cuda/cuda/cuda_config.h" to "your_python_path/site-packages/tensorflow/include/tensorflow/stream_executor/cuda/cuda_config.h".

• run the following script:

mkdir build
cd build && cmake ..
make

• run "test_op.py" and check the numeric errors to test your install

• follow the below codes snippet to integrate this Op into your own code:

  p_module = tf.load_op_library(so_lib_path)
  s_roi_align = op_module.ps_roi_align
  
  ops.RegisterGradient("PsRoiAlign")
  ef _ps_roi_align_grad(op, grad, _):
   '''The gradients for `PsRoiAlign`.
   '''
   inputs_features = op.inputs[0]
   rois = op.inputs[1]
   pooled_features_grad = op.outputs[0]
   pooled_index = op.outputs[1]
   grid_dim_width = op.get_attr('grid_dim_width')
   grid_dim_height = op.get_attr('grid_dim_height')
  
   return [op_module.ps_roi_align_grad(inputs_features, rois, grad, pooled_index, grid_dim_width, grid_dim_height, pool_method), None]
  
  ool_method = 'max' # or 'mean'
  ool_result = ps_roi_align(features, rois, 2, 2, pool_method)
  

The code is tested under TensorFlow 1.6 with CUDA 8.0 using Ubuntu 16.04. This PsRoIAlign Op had been used to train Xception based Light-Head RCNN successfully with performance at ~75%mAP on PASCAL VOC 2007 Test dataset, you can see codes here.

Update:

• Added support for mean pooling (default is max pooling)
• PsRoIAlign now support oriented RoI inputs (for both max and mean pooling).

Future Work:

• Check if there is need to ensure the convex of polygon
• Improve performance

If you encountered some linkage problem when generating or loading *.so, you are highly recommended to read this section in the official tourial to make sure you were using the same C++ ABI version.

The MIT License