Add ROIAlign pooling

[JihongJu]

Using tensorflow.image.crop_and_resize. crop_and_resize uses the bounding boxes defined by the normalized coordinates of the top left and bottom right corners [y1, x1, y2, x2] instead of [x, y, w, h].

As proposed in broadinstitute/keras-rcnn#16.

[codecov-io]

Codecov Report

Merging #22 into master will increase coverage by 3.82%. The diff coverage is 100%.

@@            Coverage Diff             @@
##           master      #22      +/-   ##
==========================================
+ Coverage   51.46%   55.29%   +3.82%     
==========================================
  Files          14       14              
  Lines         579      586       +7     
==========================================
+ Hits          298      324      +26     
+ Misses        281      262      -19

Impacted Files	Coverage Δ
keras_rcnn/layers/pooling.py	100% <100%> (+80%)	:arrow_up:
keras_rcnn/layers/__init__.py	100% <100%> (ø)	:arrow_up:
keras_rcnn/backend/tensorflow_backend.py	96.92% <100%> (+0.14%)	:arrow_up:
...ayers/object_detection/_region_proposal_network.py	30.43% <0%> (-39.14%)	:arrow_down:
keras_rcnn/losses/rpn.py	100% <0%> (ø)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 5efd6bc...237e616. Read the comment docs.

[JihongJu]

@0x00b1 Maybe review again?

[0x00b1]

@JihongJu How would you feel about replacing ROI with ROIAlign?

[JihongJu]

@0x00b1 I agree with you. According to Mask R-CNN (He et. al.), ROIAlign "improves AP by ∼3 points and AP75 by ∼5 point". And I think this is significant enough.

[0x00b1]

@JihongJu Totally. I guess you collapse ROIAlign into the ROI class.

[JihongJu]

@0x00b1 Done

[0x00b1]

LGTM