Can you explain the ROIAlignLayer?

[cicelyli]

Hi, I was trying to read the source code of ROIAlignLayer.cpp and had some difficulties understanding the bilinear interpolation part in your implementation as following.

          for (int h_idx = floor(hstart); h_idx <= ceil(hend) && h_idx < height_; ++h_idx) {
            for (int w_idx = floor(wstart); w_idx <= ceil(wend) && w_idx < width_; ++w_idx) {
              if (counter < 4) {
                b_index[counter] = ((((n*channels_ + c) * height_) + h_idx ) * width_ )+ w_idx; 
                b_index_curr[counter] = (h_idx*width_) + w_idx; 
                //Normalize h_idx and w_idx
                h_idx_n =  static_cast<Dtype>( (2*(static_cast<Dtype>(h_idx) - roi_start_h) / (roi_end_h - roi_start_h)) - 1);
                w_idx_n =  static_cast<Dtype>( (2*(static_cast<Dtype>(w_idx) - roi_start_w) / (roi_end_w - roi_start_w))  - 1);
                h_idx_n = min(max(h_idx_n, static_cast<Dtype>(-1.0)),one);
                w_idx_n = min(max(w_idx_n, static_cast<Dtype>(-1.0)),one);
                multiplier[counter] = max(zero,static_cast<Dtype>(1 - fabs(x_smp_n - w_idx_n))) 
                                         * max(zero,static_cast<Dtype>(1 - fabs(y_smp_n - h_idx_n)));

                bisampled[smp/2] += batch_data[b_index_curr[counter]]*multiplier[counter];  
                ++counter; 
	 } else { 
	    goto stop;  
	 }
            } // w_idx
          } //h_idx

Suppose if a ROI bin's width ranges from 0.5 to 2.1, and height ranges from 0.9 to 1.5. Then the hstart=0.9, hend=1.5, wstart=0.5, wend =2.1. And suppose smp = 0.

Then the loop starts with h_idx = floor(0.9) = 0, w_idx =floor(0.5) = 0, bisampled[0] is added with the feature map value at (0,0) multiplying a weight. count = 1.

In the second loop, h_idx = 0, ++w_idx =1, bisampled[0] is added with the feature map value at (0,1) multiplying a bilinear interpolation weight, count = 2

In the third loop, h_idx = 0, ++w_idx =2, bisampled[0] is added with the feature map value at (0,2) multiplying a bilinear interpolation weight. count = 3

In the fourth loop, h_idx = 0, ++w_idx =3, bisampled[0] is added with the feature map value at (0,3) multiplying a bilinear interpolation weight. count =4, then it goes to stop function.

My question is, to get the bisampled[0], isn't it to sum up the bilinear interpolation weighted feature map value at (0,0), (0,1),(1,0) and (1,1) since the upper-left corner of the ROI bin is at (0.9,0.5)? Or is there anything I misunderstood? Hope you can answer my question. Thanks!