VL_NNROIPOOL - CNN region of interest pooling.

Y = VL_NNROIPOOL(X, ROIS) pools each feature channel in X in the specified regions of interest ROIS. ROIS is a 5 x K array containing K regions. Each region has five coordinates [t, u0, v0, u1, v1] where u0, v0 is the upper-left corner of a ROI, u1, v1 is the bottom-right corner, and t is the index of the image that contains the region. Spatial coordiantes start at (1,1), with u indexing the horizontal axis and v the vertical one. The image indeces ranges from 1 to the number of images stored in the tensor X.

If X has C feature channels, then the output Y is a 1 x 1 x C x K array, with one image instance per region. Arguments can be SINGLE or DOUBLE and CPU or GPU arrays; however, they must all be of the same type (unless empty).

DZDX = VL_NNROIPOOL(X, ROIS, DZDY) computes the derivative of the layer projected on DZDY with respect to X.

VL_NNROIPOOL(___, 'opt', value, ...) accepts the following options:

• Method ['max']

  Choose between 'max' and 'avg' (average) pooling.

• Subdivisions [[1 1]]

  Specifies the number [SH,SW] of vertical and horizontal tiles of a region. This makes the output a SH x SW x C x K array.

• Transform [1]

  Specifies a spatial transformation to apply to region vertices before they are applied to the input tensor. If T is a scalar, then the transformation is a scaling centered at the origin:

     u' = T (u - 1) + 1,
     v' = T (v - 1) + 1.
  

  If T is a 2D vector, then different scaling factors for the u and v can be specified. Finally, if T is a 2 x 2 matrix, then:

     u' = T(1,1) u + T(1,2) v + T(1,3),
     v' = T(2,1) u + T(2,2) v + T(2,3).
  

  Note that only the upper-left and bottom-right corners of each rectangular region are transformed. Thus this is mostly useful for axis-aligned transformations; the generality of the expression allows, however, to swap u and v, which may be needed to match different conventions for the box coordiantes.

See also: VL_NNPOOL().