Why you use fft_shift_implement for calculating 3D IOU

[monroyaume5]

def iou3d(box_xyzwhd_1, box_xyzwhd_2, input_size):
    """ Numpy version of 3D bounding box IOU calculation 
    Args:
        box_xyzwhd_1        ->      box1 [x, y, z, w, h, d]
        box_xyzwhd_2        ->      box2 [x, y, z, w, h, d]"""
    assert box_xyzwhd_1.shape[-1] == 6
    assert box_xyzwhd_2.shape[-1] == 6
    fft_shift_implement = np.array([0, 0, input_size[2]/2])
    ### areas of both boxes
    box1_area = box_xyzwhd_1[..., 3] * box_xyzwhd_1[..., 4] * box_xyzwhd_1[..., 5]
    box2_area = box_xyzwhd_2[..., 3] * box_xyzwhd_2[..., 4] * box_xyzwhd_2[..., 5]
    ### find the intersection box
    box1_min = box_xyzwhd_1[..., :3] + fft_shift_implement - box_xyzwhd_1[..., 3:] * 0.5
    box1_max = box_xyzwhd_1[..., :3] + fft_shift_implement + box_xyzwhd_1[..., 3:] * 0.5
    box2_min = box_xyzwhd_2[..., :3] + fft_shift_implement - box_xyzwhd_2[..., 3:] * 0.5
    box2_max = box_xyzwhd_2[..., :3] + fft_shift_implement + box_xyzwhd_2[..., 3:] * 0.5

    # box1_min = box_xyzwhd_1[..., :3] - box_xyzwhd_1[..., 3:] * 0.5
    # box1_max = box_xyzwhd_1[..., :3] + box_xyzwhd_1[..., 3:] * 0.5
    # box2_min = box_xyzwhd_2[..., :3] - box_xyzwhd_2[..., 3:] * 0.5
    # box2_max = box_xyzwhd_2[..., :3] + box_xyzwhd_2[..., 3:] * 0.5

    left_top = np.maximum(box1_min, box2_min)
    bottom_right = np.minimum(box1_max, box2_max)
    ### get intersection area
    intersection = np.maximum(bottom_right - left_top, 0.0)
    intersection_area = intersection[..., 0] * intersection[..., 1] * intersection[..., 2]
    ### get union area
    union_area = box1_area + box2_area - intersection_area
    ### get iou
    iou = np.nan_to_num(intersection_area / (union_area + 1e-10))
    return iou

Could you please explain why you are using fft_shift_implement for the calculation of 3D IOU? It doesn't seem to have any effect on the final value of the IOU.

[ZhangAoCanada]

Hi,

To explain your question: Our radar configurations were set as maximum 64 bins on Doppler axis. This means, whenever the velocity of an object exceeds the maximum detection speed, the RAD spectrum of this object will jump from the left to the right on Doppler Axis. The shifting operation is to tackle this issue.