def grid_cell_to_tile_coords()

in 04_detect_segment/utils_box.py [0:0]

• 19 lines of code
• 5 McCabe index (conditional complexity)

def grid_cell_to_tile_coords(rois, grid_n, tile_size):
    # converts between coordinates used internally by the model
    # and coordinates expected by Tensorflow's draw_bounding_boxes function
    #
    # input coords:
    # shape [batch, grid_n, grid_n, n, 3]
    # coordinates in last dimension are x, y, w
    # x and y are in [-1, 1] relative to grid cell center and size of grid cell
    # w is in [0, 1] relatively to tile size. w is a "diameter", not "radius"
    #
    # output coords:
    # shape [batch, grid_n, grid_n, n, 4]
    # coordinates in last dimension are y1, x1, y2, x2
    # relatively to tile_size

    # grid for (0,0) based tile of size tile_size
    cell_w = tile_size/grid_n
    grid = gen_grid(grid_n) * cell_w
    # grid cell centers
    grid_centers = (grid + grid + cell_w) / 2.0  # shape [grid_n, grid_n, 2]
    # roi coordinates
    roi_cx, roi_cy, roi_w = tf.unstack(rois, axis=-1) # shape [batch, grid_n, grid_n, n]
    # grid centers unstacked
    gr_cx, gr_cy = tf.unstack(grid_centers, axis=-1) # shape [grid_n, grid_n]
    gr_cx = tf.expand_dims(tf.expand_dims(gr_cx, 0), 3) # shape [1, grid_n, grid_n, 1]
    gr_cy = tf.expand_dims(tf.expand_dims(gr_cy, 0), 3) # shape [1, grid_n, grid_n, 1]
    roi_cx = roi_cx * cell_w/2 # roi_x=1 means cell center + cell_w/2
    roi_cx = roi_cx+gr_cx
    roi_cy = roi_cy * cell_w/2 # roi_x=1 means cell center + cell_w/2
    roi_cy = roi_cy+gr_cy
    roi_w = roi_w * tile_size
    roi_x1 = roi_cx - roi_w/2
    roi_x2 = roi_cx + roi_w/2
    roi_y1 = roi_cy - roi_w/2
    roi_y2 = roi_cy + roi_w/2
    rois = tf.stack([roi_x1, roi_y1, roi_x2, roi_y2], axis=4)  # shape [batch, grid_n, grid_n, n, 4]
    return rois