Hi, argman, I have some doubt about EAST. I need your help here. 1、Do you have train with focal loss， Does it work for improve the performance?

I have write a focal loss code, can you help me to check this implement is right or wrong?

def focal_loss(y_true_cls, y_pred_cls, training_mask, normalizer, gamma = 2, alpha =0.25): epsilon = 10e-8 y_pred_cls = tf.clip_by_value(y_pred_cls, epsilon, 1 - epsilon) logits = tf.log(y_pred_cls / (1. - y_pred_cls)) per_entry_cross_ent = (tf.nn.sigmoid_cross_entropy_with_logits( labels=y_true_cls, logits=logits)) prediction_probabilities = y_pred_cls p_t = ((y_true_cls prediction_probabilities) + ((1 - y_true_cls) (1 - prediction_probabilities))) modulating_factor = 1.0 if gamma: modulating_factor = tf.pow(1.0 - p_t, gamma) alpha_weight_factor = 1.0 if alpha is not None: alpha_weight_factor = (y_true_cls alpha + (1 - y_true_cls) (1 - alpha)) focal_cross_entropy_loss = (modulating_factor alpha_weight_factor per_entry_cross_ent) total_loss = tf.reduce_sum(focal_cross_entropy_loss * training_mask) total_loss /= normalizer tf.summary.scalar('classification_focal_loss', total_loss) return total_loss

def loss(y_true_cls, y_pred_cls, y_true_geo, y_pred_geo, training_mask): normalizer = tf.reduce_sum(tf.cast(tf.equal(y_true_cls * training_mask, 1), tf.int32)) + 1 normalizer = tf.cast(normalizer, tf.float32) tf.summary.scalar('normalizer', normalizer) classification_loss = focal_loss(y_true_cls, y_pred_cls, training_mask, normalizer)

d1 -> top, d2->right, d3->bottom, d4->left

d1_gt, d2_gt, d3_gt, d4_gt, theta_gt = tf.split(value=y_true_geo, num_or_size_splits=5, axis=3)
d1_pred, d2_pred, d3_pred, d4_pred, theta_pred = tf.split(value=y_pred_geo, num_or_size_splits=5, axis=3)
area_gt = (d1_gt + d3_gt) * (d2_gt + d4_gt)
area_pred = (d1_pred + d3_pred) * (d2_pred + d4_pred)
w_union = tf.minimum(d2_gt, d2_pred) + tf.minimum(d4_gt, d4_pred)
h_union = tf.minimum(d1_gt, d1_pred) + tf.minimum(d3_gt, d3_pred)
area_intersect = w_union * h_union
area_union = area_gt + area_pred - area_intersect
L_AABB = -tf.log((area_intersect + 1.0)/(area_union + 1.0))
L_theta = 1 - tf.cos(theta_pred - theta_gt)
tf.summary.scalar('geometry_AABB', tf.reduce_sum(L_AABB * y_true_cls * training_mask) / normalizer)
tf.summary.scalar('geometry_theta', tf.reduce_sum(L_theta * y_true_cls * training_mask) / normalizer)
L_g = L_AABB + 20 * L_theta
return tf.reduce_sum(L_g * y_true_cls * training_mask) / normalizer + classification_loss

2、how to choose score threshold and nms threshold ? Which score threshold and nms threshold will lead the Hmean best?

argman / EAST

is focal loss help to imporve performance? how to decide score threshold when in test? #157

d1 -> top, d2->right, d3->bottom, d4->left