retina_model/anchor.py

"""Anchor utils modified from https://github.com/biubug6/Pytorch_Retinaface"""
import math
import tensorflow as tf
import numpy as np
from itertools import product as product


###############################################################################
#   Tensorflow / Numpy Priors                                                 #
###############################################################################
def prior_box(image_sizes, min_sizes, steps, clip=False):
    """prior box"""
    feature_maps = [
        [math.ceil(image_sizes[0] / step), math.ceil(image_sizes[1] / step)]
        for step in steps]

    anchors = []
    for k, f in enumerate(feature_maps):
        for i, j in product(range(f[0]), range(f[1])):
            for min_size in min_sizes[k]:
                s_kx = min_size / image_sizes[1]
                s_ky = min_size / image_sizes[0]
                cx = (j + 0.5) * steps[k] / image_sizes[1]
                cy = (i + 0.5) * steps[k] / image_sizes[0]
                anchors += [cx, cy, s_kx, s_ky]

    output = np.asarray(anchors).reshape([-1, 4])

    if clip:
        output = np.clip(output, 0, 1)

    return output


def prior_box_tf(image_sizes, min_sizes, steps, clip=False):
    """prior box"""
    image_sizes = tf.cast(tf.convert_to_tensor(image_sizes), tf.float32)
    feature_maps = tf.math.ceil(
        tf.reshape(image_sizes, [1, 2]) /
        tf.reshape(tf.cast(steps, tf.float32), [-1, 1]))

    anchors = []
    for k in range(len(min_sizes)):
        grid_x, grid_y = _meshgrid_tf(tf.range(feature_maps[k][1]),
                                      tf.range(feature_maps[k][0]))
        cx = (grid_x + 0.5) * steps[k] / image_sizes[1]
        cy = (grid_y + 0.5) * steps[k] / image_sizes[0]
        cxcy = tf.stack([cx, cy], axis=-1)
        cxcy = tf.reshape(cxcy, [-1, 2])
        cxcy = tf.repeat(cxcy, repeats=tf.shape(min_sizes[k])[0], axis=0)

        sx = min_sizes[k] / image_sizes[1]
        sy = min_sizes[k] / image_sizes[0]
        sxsy = tf.stack([sx, sy], 1)
        sxsy = tf.repeat(sxsy[tf.newaxis],
                         repeats=tf.shape(grid_x)[0] * tf.shape(grid_x)[1],
                         axis=0)
        sxsy = tf.reshape(sxsy, [-1, 2])

        anchors.append(tf.concat([cxcy, sxsy], 1))

    output = tf.concat(anchors, axis=0)

    if clip:
        output = tf.clip_by_value(output, 0, 1)

    return output


def _meshgrid_tf(x, y):
    """ workaround solution of the tf.meshgrid() issue:
        https://github.com/tensorflow/tensorflow/issues/34470"""
    grid_shape = [tf.shape(y)[0], tf.shape(x)[0]]
    grid_x = tf.broadcast_to(tf.reshape(x, [1, -1]), grid_shape)
    grid_y = tf.broadcast_to(tf.reshape(y, [-1, 1]), grid_shape)
    return grid_x, grid_y


###############################################################################
#   Tensorflow Encoding                                                       #
###############################################################################
def encode_tf(labels, priors, match_thresh, ignore_thresh,
              variances=[0.1, 0.2]):
    """tensorflow encoding"""
    assert ignore_thresh <= match_thresh
    priors = tf.cast(priors, tf.float32)
    bbox = labels[:, :4]
    landm = labels[:, 4:-1]
    landm_valid = labels[:, -1]  # 1: with landm, 0: w/o landm.

    # jaccard index
    overlaps = _jaccard(bbox, _point_form(priors))

    # (Bipartite Matching)
    # [num_objects] best prior for each ground truth
    best_prior_overlap, best_prior_idx = tf.math.top_k(overlaps, k=1)
    best_prior_overlap = best_prior_overlap[:, 0]
    best_prior_idx = best_prior_idx[:, 0]

    # [num_priors] best ground truth for each prior
    overlaps_t = tf.transpose(overlaps)
    best_truth_overlap, best_truth_idx = tf.math.top_k(overlaps_t, k=1)
    best_truth_overlap = best_truth_overlap[:, 0]
    best_truth_idx = best_truth_idx[:, 0]

    # ensure best prior
    def _loop_body(i, bt_idx, bt_overlap):
        bp_mask = tf.one_hot(best_prior_idx[i], tf.shape(bt_idx)[0])
        bp_mask_int = tf.cast(bp_mask, tf.int32)
        new_bt_idx = bt_idx * (1 - bp_mask_int) + bp_mask_int * i
        bp_mask_float = tf.cast(bp_mask, tf.float32)
        new_bt_overlap = bt_overlap * (1 - bp_mask_float) + bp_mask_float * 2
        return tf.cond(best_prior_overlap[i] > match_thresh,
                       lambda: (i + 1, new_bt_idx, new_bt_overlap),
                       lambda: (i + 1, bt_idx, bt_overlap))
    _, best_truth_idx, best_truth_overlap = tf.while_loop(
        lambda i, bt_idx, bt_overlap: tf.less(i, tf.shape(best_prior_idx)[0]),
        _loop_body, [tf.constant(0), best_truth_idx, best_truth_overlap])

    matches_bbox = tf.gather(bbox, best_truth_idx)  # [num_priors, 4]
    matches_landm = tf.gather(landm, best_truth_idx)  # [num_priors, 10]
    matches_landm_v = tf.gather(landm_valid, best_truth_idx)  # [num_priors]

    loc_t = _encode_bbox(matches_bbox, priors, variances)
    landm_t = _encode_landm(matches_landm, priors, variances)
    landm_valid_t = tf.cast(matches_landm_v > 0, tf.float32)
    conf_t = tf.cast(best_truth_overlap > match_thresh, tf.float32)
    conf_t = tf.where(
        tf.logical_and(best_truth_overlap < match_thresh,
                       best_truth_overlap > ignore_thresh),
        tf.ones_like(conf_t) * -1, conf_t)    # 1: pos, 0: neg, -1: ignore

    return tf.concat([loc_t, landm_t, landm_valid_t[..., tf.newaxis],
                      conf_t[..., tf.newaxis]], axis=1)


def _encode_bbox(matched, priors, variances):
    """Encode the variances from the priorbox layers into the ground truth
    boxes we have matched (based on jaccard overlap) with the prior boxes.
    Args:
        matched: (tensor) Coords of ground truth for each prior in point-form
            Shape: [num_priors, 4].
        priors: (tensor) Prior boxes in center-offset form
            Shape: [num_priors,4].
        variances: (list[float]) Variances of priorboxes
    Return:
        encoded boxes (tensor), Shape: [num_priors, 4]
    """

    # dist b/t match center and prior's center
    g_cxcy = (matched[:, :2] + matched[:, 2:]) / 2 - priors[:, :2]
    # encode variance
    g_cxcy /= (variances[0] * priors[:, 2:])
    # match wh / prior wh
    g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
    g_wh = tf.math.log(g_wh) / variances[1]
    # return target for smooth_l1_loss
    return tf.concat([g_cxcy, g_wh], 1)  # [num_priors,4]


def _encode_landm(matched, priors, variances):
    """Encode the variances from the priorbox layers into the ground truth
    boxes we have matched (based on jaccard overlap) with the prior boxes.
    Args:
        matched: (tensor) Coords of ground truth for each prior in point-form
            Shape: [num_priors, 10].
        priors: (tensor) Prior boxes in center-offset form
            Shape: [num_priors,4].
        variances: (list[float]) Variances of priorboxes
    Return:
        encoded landm (tensor), Shape: [num_priors, 10]
    """

    # dist b/t match center and prior's center
    matched = tf.reshape(matched, [tf.shape(matched)[0], 5, 2])
    priors = tf.broadcast_to(
        tf.expand_dims(priors, 1), [tf.shape(matched)[0], 5, 4])
    g_cxcy = matched[:, :, :2] - priors[:, :, :2]
    # encode variance
    g_cxcy /= (variances[0] * priors[:, :, 2:])
    # g_cxcy /= priors[:, :, 2:]
    g_cxcy = tf.reshape(g_cxcy, [tf.shape(g_cxcy)[0], -1])
    # return target for smooth_l1_loss
    return g_cxcy


def _point_form(boxes):
    """ Convert prior_boxes to (xmin, ymin, xmax, ymax)
    representation for comparison to point form ground truth data.
    Args:
        boxes: (tensor) center-size default boxes from priorbox layers.
    Return:
        boxes: (tensor) Converted xmin, ymin, xmax, ymax form of boxes.
    """
    return tf.concat((boxes[:, :2] - boxes[:, 2:] / 2,
                      boxes[:, :2] + boxes[:, 2:] / 2), axis=1)


def _intersect(box_a, box_b):
    """ We resize both tensors to [A,B,2]:
    [A,2] -> [A,1,2] -> [A,B,2]
    [B,2] -> [1,B,2] -> [A,B,2]
    Then we compute the area of intersect between box_a and box_b.
    Args:
      box_a: (tensor) bounding boxes, Shape: [A,4].
      box_b: (tensor) bounding boxes, Shape: [B,4].
    Return:
      (tensor) intersection area, Shape: [A,B].
    """
    A = tf.shape(box_a)[0]
    B = tf.shape(box_b)[0]
    max_xy = tf.minimum(
        tf.broadcast_to(tf.expand_dims(box_a[:, 2:], 1), [A, B, 2]),
        tf.broadcast_to(tf.expand_dims(box_b[:, 2:], 0), [A, B, 2]))
    min_xy = tf.maximum(
        tf.broadcast_to(tf.expand_dims(box_a[:, :2], 1), [A, B, 2]),
        tf.broadcast_to(tf.expand_dims(box_b[:, :2], 0), [A, B, 2]))
    inter = tf.maximum((max_xy - min_xy), tf.zeros_like(max_xy - min_xy))
    return inter[:, :, 0] * inter[:, :, 1]


def _jaccard(box_a, box_b):
    """Compute the jaccard overlap of two sets of boxes.  The jaccard overlap
    is simply the intersection over union of two boxes.  Here we operate on
    ground truth boxes and default boxes.
    E.g.:
        A ∩ B / A ∪ B = A ∩ B / (area(A) + area(B) - A ∩ B)
    Args:
        box_a: (tensor) Ground truth bounding boxes, Shape: [num_objects,4]
        box_b: (tensor) Prior boxes from priorbox layers, Shape: [num_priors,4]
    Return:
        jaccard overlap: (tensor) Shape: [box_a.size(0), box_b.size(0)]
    """
    inter = _intersect(box_a, box_b)
    area_a = tf.broadcast_to(
        tf.expand_dims(
            (box_a[:, 2] - box_a[:, 0]) * (box_a[:, 3] - box_a[:, 1]), 1),
        tf.shape(inter))  # [A,B]
    area_b = tf.broadcast_to(
        tf.expand_dims(
            (box_b[:, 2] - box_b[:, 0]) * (box_b[:, 3] - box_b[:, 1]), 0),
        tf.shape(inter))  # [A,B]
    union = area_a + area_b - inter
    return inter / union  # [A,B]


###############################################################################
#   Tensorflow Decoding                                                       #
###############################################################################
def decode_tf(labels, priors, variances=[0.1, 0.2]):
    """tensorflow decoding"""
    bbox = _decode_bbox(labels[:, :4], priors, variances)
    landm = _decode_landm(labels[:, 4:14], priors, variances)
    landm_valid = labels[:, 14][:, tf.newaxis]
    conf = labels[:, 15][:, tf.newaxis]

    return tf.concat([bbox, landm, landm_valid, conf], axis=1)


def _decode_bbox(pre, priors, variances=[0.1, 0.2]):
    """Decode locations from predictions using priors to undo
    the encoding we did for offset regression at train time.
    Args:
        pre (tensor): location predictions for loc layers,
            Shape: [num_priors,4]
        priors (tensor): Prior boxes in center-offset form.
            Shape: [num_priors,4].
        variances: (list[float]) Variances of priorboxes
    Return:
        decoded bounding box predictions
    """
    centers = priors[:, :2] + pre[:, :2] * variances[0] * priors[:, 2:]
    sides = priors[:, 2:] * tf.math.exp(pre[:, 2:] * variances[1])

    return tf.concat([centers - sides / 2, centers + sides / 2], axis=1)


def _decode_landm(pre, priors, variances=[0.1, 0.2]):
    """Decode landm from predictions using priors to undo
    the encoding we did for offset regression at train time.
    Args:
        pre (tensor): landm predictions for loc layers,
            Shape: [num_priors,10]
        priors (tensor): Prior boxes in center-offset form.
            Shape: [num_priors,4].
        variances: (list[float]) Variances of priorboxes
    Return:
        decoded landm predictions
    """
    landms = tf.concat(
        [priors[:, :2] + pre[:, :2] * variances[0] * priors[:, 2:],
         priors[:, :2] + pre[:, 2:4] * variances[0] * priors[:, 2:],
         priors[:, :2] + pre[:, 4:6] * variances[0] * priors[:, 2:],
         priors[:, :2] + pre[:, 6:8] * variances[0] * priors[:, 2:],
         priors[:, :2] + pre[:, 8:10] * variances[0] * priors[:, 2:]], axis=1)
    return landms