def peaks(boxes: np.ndarray, samples_per_box: int) -> Mapping[int, Iterable[int]]:
boxes = boxes.transpose((2, 0, 1, 3))
horiz_bin_count, vert_bin_count, box_width, box_height = boxes.shape
boxes = boxes.reshape(boxes.shape[:2] + (-1,))
indices_peaks = boxes.argpartition(- samples_per_box)[..., - samples_per_box:]
# We compute the frequency and sample indices for the peaks.
indices_freqs = np.arange(horiz_bin_count)[:, np.newaxis, np.newaxis] * box_height + indices_peaks % box_height
indices_samples = np.arange(vert_bin_count)[np.newaxis, :, np.newaxis] * box_width + indices_peaks // box_height
freqs_dict = defaultdict(list)
for freq_idx, sample_idx in np.nditer([indices_freqs, indices_samples]):
freqs_dict[freq_idx.item()].append(sample_idx.item())
return freqs_dict
