How to change the RROI 's calculation process

[lllukehuang]

My current work needs to change the feature extraction method of RROI layer to obtain better image context. Specifically, I need to expand the width of the ROI region output by RPN, but I'm not sure where to make the change.

My previous attempts: (in layers/roi_align_rotated.py)

    def forward(self, input, rois):
        """
        Args:
            input: NCHW images
            rois: Bx6 boxes. First column is the index into N.
                The other 5 columns are (x_ctr, y_ctr, width, height, angle_degrees).
        """
        assert rois.dim() == 2 and rois.size(1) == 6
        orig_dtype = input.dtype
        if orig_dtype == torch.float16:
            input = input.float()
            rois = rois.float()
        return roi_align_rotated(
            input, rois, self.output_size, self.spatial_scale, self.sampling_ratio
        ).to(dtype=orig_dtype)

Just simply expand the width of rois here:

    def forward(self, input, rois):
        """
        Args:
            input: NCHW images
            rois: Bx6 boxes. First column is the index into N.
                The other 5 columns are (x_ctr, y_ctr, width, height, angle_degrees).
        """
        assert rois.dim() == 2 and rois.size(1) == 6
        orig_dtype = input.dtype
        if orig_dtype == torch.float16:
            input = input.float()
            rois = rois.float()
        for box in rois:
            box[3] = box[3] * 2
        return roi_align_rotated(
            input, rois, self.output_size, self.spatial_scale, self.sampling_ratio
        ).to(dtype=orig_dtype)

But actually, in this way the model's inference results just looks same as before.

So my questions are:

• where is the RROI implementation part of codes?
• Am I making the right change on expanding the feature for the feature map?

Any suggestion is welcome.