Closed zhangxuan1918 closed 4 years ago
we notice over-fitting during training, to improve, we randomly argument input data
tx = random(0, 32) ty = random(0, 32) render_image_size = 224 image_resized = tf.image.resize(image, (256, 256)) image_shift = tf.image.crop_to_bounding_box(image_resized, tx, ty, render_image_size, render_image_size) # update landmark gt lm = mat_data['pt2d'] lm_shift = np.copy(lm) lm_shift *= 256. / 450. lm_shift[0], lm_shift[1] = lm_shift[0] - ty, lm_shift[1] - tx lm *= 224. / 450 # update pose params pp = mat_data['Pose_Para'] pp[0, 3:5] = pp[0, 3:5] * 256 / 450 pp[0, 6] = pp[0, 6] * 256 / 450 pp[0, 3], pp[0, 4] = pp[0, 3] - ty, pp[0, 4] - (32 - tx)
we notice over-fitting during training, to improve, we randomly argument input data