import cv2 import numpy as np import matplotlib import math import torch import random import copy from dwpose_utils.dwpose_detector import build_pose_dict from torch.nn import functional as F eps = 0.01 DROP_FACE_POINTS = {0, 14, 15, 16, 17} DROP_UPPER_POINTS = {0, 14, 15, 16, 17, 2, 1, 5, 3, 6} DROP_LOWER_POINTS = {8, 9, 10, 11, 12, 13} def generate_pose_plan(pose_aug_probs): p1, p2, p3, p4 = pose_aug_probs r = random.random() if r < p1: return {"mode": "drop_point", "point_idx": random.randint(0, 17)} elif r < p1 + p2: return {"mode": "stretch_limb", "limb_idx": random.randint(0, 18), "stretch_scale": random.uniform(0.5, 2.0)} elif r < p1 + p2 + p3: region_mode = random.choice(["face", "upper", "lower"]) if region_mode == "face": pts = DROP_FACE_POINTS elif region_mode == "upper": pts = DROP_UPPER_POINTS else: pts = DROP_LOWER_POINTS return {"mode": "drop_region", "region": region_mode, "points": pts} else: return {"mode": "none"} def get_stickwidth(W, H, stickwidth=4): if max(W, H) < 512: ratio = 1.0 elif max(W, H) < 1080: ratio = 1.5 elif max(W, H) < 2160: ratio = 2.0 elif max(W, H) < 3240: ratio = 2.5 elif max(W, H) < 4320: ratio = 3.5 elif max(W, H) < 5400: ratio = 4.5 else: ratio = 4.0 return int(stickwidth * ratio) def alpha_blend_color(color, alpha): """blend color according to point conf """ return [int(c * alpha) for c in color] def draw_bodypose_aligned(canvas, candidate, subset, score, plan=None, # 新增，用于全视频固定增强 random_drop=False): H, W, C = canvas.shape candidate = np.array(candidate) subset = np.array(subset) stickwidth = get_stickwidth(W, H, stickwidth=3) limbSeq = [ [2, 3], [2, 6], [3, 4], [4, 5], [6, 7], [7, 8], [2, 9], [9, 10], [10, 11], [2, 12], [12, 13], [13, 14], [2, 1], [1, 15], [15, 17], [1, 16], [16, 18], [3, 17], [6, 18]] colors = [ [255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0], [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255], [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]] HIDE_JOINTS = set() stretch_limb_idx = None stretch_scale = None if plan: if plan["mode"] == "drop_point": HIDE_JOINTS.add(plan["point_idx"]) elif plan["mode"] == "drop_region": HIDE_JOINTS |= set(plan["points"]) elif plan["mode"] == "stretch_limb": stretch_limb_idx = plan["limb_idx"] stretch_scale = plan["stretch_scale"] hide_joint = np.zeros_like(subset, dtype=bool) for i in range(17): for n in range(len(subset)): idx_pair = limbSeq[i] if any(j in HIDE_JOINTS for j in idx_pair): continue index = subset[n][np.array(idx_pair) - 1] conf = score[n][np.array(idx_pair) - 1] if -1 in index: continue # color lighten alpha = max(conf[0] * conf[1], 0) if conf[0]>0 and conf[1]>0 else 0.35 Y = candidate[index.astype(int), 0] * float(W) X = candidate[index.astype(int), 1] * float(H) if stretch_limb_idx == i: vec_x = X[1] - X[0] vec_y = Y[1] - Y[0] X[1] = X[0] + vec_x * stretch_scale Y[1] = Y[0] + vec_y * stretch_scale hide_joint[n, idx_pair[1]-1] = True mX = np.mean(X) mY = np.mean(Y) length = ((X[0]-X[1])**2 + (Y[0]-Y[1])**2) ** 0.5 angle = math.degrees(math.atan2(X[0]-X[1], Y[0]-Y[1])) polygon = cv2.ellipse2Poly((int(mY), int(mX)), (int(length/2), stickwidth), int(angle), 0, 360, 1) cv2.fillConvexPoly(canvas, polygon, alpha_blend_color(colors[i], alpha)) canvas = (canvas * 0.6).astype(np.uint8) for i in range(18): if i in HIDE_JOINTS: continue for n in range(len(subset)): if hide_joint[n, i]: continue index = int(subset[n][i]) if index == -1: continue x, y = candidate[index][0:2] conf = score[n][i] # head_change=True alpha = 0 if conf==-2 else max(conf, 0) x = int(x * W) y = int(y * H) cv2.circle(canvas, (x, y), stickwidth, alpha_blend_color(colors[i], alpha), thickness=-1) return canvas def draw_handpose_aligned(canvas, all_hand_peaks, all_hand_scores, draw_th=0.3,): H, W, C = canvas.shape stickwidth = get_stickwidth(W, H, stickwidth=2) line_thickness = get_stickwidth(W, H, stickwidth=2) edges = [[0, 1], [1, 2], [2, 3], [3, 4], [0, 5], [5, 6], [6, 7], [7, 8], [0, 9], [9, 10], \ [10, 11], [11, 12], [0, 13], [13, 14], [14, 15], [15, 16], [0, 17], [17, 18], [18, 19], [19, 20]] for peaks, scores in zip(all_hand_peaks, all_hand_scores): for ie, e in enumerate(edges): if scores[e[0]] < draw_th or scores[e[1]] < draw_th: continue x1, y1 = peaks[e[0]] x2, y2 = peaks[e[1]] x1 = int(x1 * W) y1 = int(y1 * H) x2 = int(x2 * W) y2 = int(y2 * H) score = int(scores[e[0]] * scores[e[1]] * 255) if x1 > eps and y1 > eps and x2 > eps and y2 > eps: # 原来是2 cv2.line(canvas, (x1, y1), (x2, y2), matplotlib.colors.hsv_to_rgb([ie / float(len(edges)), 1.0, 1.0]) * score, thickness=line_thickness) for i, keyponit in enumerate(peaks): if scores[i] < draw_th: continue x, y = keyponit x = int(x * W) y = int(y * H) score = int(scores[i] * 255) if x > eps and y > eps: cv2.circle(canvas, (x, y), stickwidth, (0, 0, score), thickness=-1) return canvas def draw_facepose_aligned(canvas, all_lmks, all_scores, draw_th=0.3,face_change=False): H, W, C = canvas.shape stickwidth = get_stickwidth(W, H, stickwidth=2) SKIP_IDX = set(range(0, 17)) SKIP_IDX |= set(range(27, 36)) for lmks, scores in zip(all_lmks, all_scores): for idx, (lmk, score) in enumerate(zip(lmks, scores)): # 跳过下巴点 if idx in SKIP_IDX: continue if score < draw_th: continue x, y = lmk x = int(x * W) y = int(y * H) conf = int(score * 255) # color lighten if face_change: conf = int(conf * 0.35) if x > eps and y > eps: cv2.circle(canvas, (x, y), stickwidth, (conf, conf, conf), thickness=-1) return canvas def draw_pose_aligned(pose, H, W, ref_w=2160, without_face=False, pose_plan=None, head_change=False, face_change=False): bodies = pose['bodies'] faces = pose['faces'] hands = pose['hands'] candidate = bodies['candidate'] subset = bodies['subset'] body_score = bodies['score'].copy() # control color if head_change: target_joints = [0, 14, 15, 16, 17] body_score[:, target_joints] = -2 sz = min(H, W) sr = (ref_w / sz) if sz != ref_w else 1 canvas = np.zeros(shape=(int(H*sr), int(W*sr), 3), dtype=np.uint8) canvas = draw_bodypose_aligned(canvas, candidate, subset, score=body_score, plan=pose_plan,) # drop上半身 → 手不要画 if pose_plan and pose_plan["mode"] == "drop_region" and pose_plan.get("region") == "upper": pass else: canvas = draw_handpose_aligned(canvas, hands, pose['hands_score']) # drop脸 → 不画脸 if pose_plan and pose_plan["mode"] == "drop_region" and pose_plan.get("region") in {"upper","face"}: pass elif not without_face: canvas = draw_facepose_aligned(canvas, faces, pose['faces_score'],face_change=face_change) # print(W,H) return cv2.resize(canvas, (W, H)) # 训练中并没有用到 def generate_pose_plan(pose_aug_probs): p1, p2, p3, p4 = pose_aug_probs r = random.random() if r < p1: return {"mode": "drop_point", "point_idx": random.randint(0, 17)} elif r < p1 + p2: return {"mode": "stretch_limb", "limb_idx": random.randint(0, 18), "stretch_scale": random.uniform(0.5, 2.0)} elif r < p1 + p2 + p3: region_mode = random.choice(["face", "upper", "lower"]) if region_mode == "face": pts = DROP_FACE_POINTS elif region_mode == "upper": pts = DROP_UPPER_POINTS else: pts = DROP_LOWER_POINTS return {"mode": "drop_region", "region": region_mode, "points": pts} else: return {"mode": "none"} def get_batch_pose(pose_data, train_indices, H, W, without_face=False, pose_aug_flag=False, head_change=False, face_change=False): if pose_aug_flag: pose_plan = generate_pose_plan((0.1, 0.1, 0.3, 0.5)) else: pose_plan = None pose_batch = [] for i in train_indices: frame_pose_img = draw_pose_aligned(pose_data[i], H, W, without_face=without_face,pose_plan=pose_plan,head_change=head_change,face_change=face_change) pose_batch.append(frame_pose_img) return np.array(pose_batch) def select_refimg_index(refimg_indices, train_indices): valid_indices = [idx for idx in refimg_indices if idx < train_indices[0] or idx > train_indices[-1]] if not valid_indices: valid_indices = refimg_indices return np.random.choice(valid_indices) def draw_face_mask(all_face_bbox, train_indices, real_t, ori_height, ori_width, for_vae = True): vae_t = 1 + (real_t-1) // 4 if len(all_face_bbox) == 0: if for_vae: return torch.ones((vae_t, ori_height, ori_width), dtype=torch.float32) else: return torch.ones((real_t, ori_height, ori_width), dtype=torch.float32) # 如果存在facebbox if for_vae: mask = torch.zeros((vae_t, ori_height, ori_width), dtype=torch.float32) else: mask = torch.zeros((real_t, ori_height, ori_width), dtype=torch.float32) for i, idx in enumerate(train_indices): bbox = all_face_bbox[idx] # 获取对应的 bbox if len(bbox) == 0: continue x1, y1, x2, y2 = map(int,bbox) if for_vae: compressed_idx = (i + 3) // 4 # 每 4 帧压缩成一帧 # 并集 mask[compressed_idx, y1:y2, x1:x2] = 1 else: mask[i, y1:y2, x1:x2] = 1 return mask def draw_face_mask_frompose(pose_data, train_indices, real_t, ori_height, ori_width, for_vae=True, # 是否 4→1 压缩 score_thresh=0.8, # 关键点平均置信度阈值 expand_ratio=0.1): # 人脸 box 适当放大一点 vae_t = 1 + (real_t - 1) // 4 if for_vae: mask = torch.zeros((vae_t, ori_height, ori_width), dtype=torch.float32) valid_flags = [True] * vae_t else: mask = torch.zeros((real_t, ori_height, ori_width), dtype=torch.float32) for i, idx in enumerate(train_indices): pose = pose_data[idx] faces = pose.get('faces', pose.get('face', [])) scores = pose.get('faces_score', pose.get('face_score', [])) if len(scores) == 0 or len(faces) == 0: if for_vae: valid_flags[(i + 3) // 4] = False continue frame_valid = True frame_mask = torch.zeros((ori_height, ori_width), dtype=torch.float32) for peaks, sc in zip(faces, scores): if np.mean(sc) < score_thresh: frame_valid = False break xys = np.array(peaks) if xys.size == 0: continue x1, y1 = int(np.min(xys[:, 0] * ori_width)), int(np.min(xys[:, 1] * ori_height)) x2, y2 = int(np.max(xys[:, 0] * ori_width)), int(np.max(xys[:, 1] * ori_height)) # 额外抬高顶部 extra_top = int((y2 - y1) * 0.4) y1 = max(0, y1 - extra_top) x1, y1, x2, y2 = expand_bbox( x1, y1, x2, y2, big=expand_ratio, small=expand_ratio, frame_width=ori_width, frame_height=ori_height ) frame_mask[y1:y2, x1:x2] = 1 if not frame_valid: if for_vae: valid_flags[(i + 3) // 4] = False continue if for_vae: compressed_idx = (i + 3) // 4 mask[compressed_idx] = torch.maximum(mask[compressed_idx], frame_mask) else: mask[i] = frame_mask if for_vae: for t, valid in enumerate(valid_flags): if not valid: mask[t] = 0 return mask def draw_hand_mask(pose_data, train_indices, real_t, ori_height, ori_width, for_vae=True, score_thresh=0.8, expand_ratio=0.2): vae_t = 1 + (real_t - 1) // 4 if for_vae: mask = torch.zeros((vae_t, ori_height, ori_width), dtype=torch.float32) valid_flags = [True] * vae_t else: mask = torch.zeros((real_t, ori_height, ori_width), dtype=torch.float32) for i, idx in enumerate(train_indices): pose = pose_data[idx] hands = pose.get('hands', []) hand_scores = pose.get('hands_score', []) if len(hand_scores) == 0 or len(hands) == 0: if for_vae: valid_flags[(i + 3) // 4] = False continue frame_valid = True # 用一个临时mask存每只手的mask，最后叠加 frame_mask = torch.zeros((ori_height, ori_width), dtype=torch.float32) for peaks, scores in zip(hands, hand_scores): if np.mean(scores) < score_thresh: frame_valid = False break xys = np.array(peaks) if xys.size == 0: continue x1, y1 = int(np.min(xys[:, 0] * ori_width)), int(np.min(xys[:, 1] * ori_height)) x2, y2 = int(np.max(xys[:, 0] * ori_width)), int(np.max(xys[:, 1] * ori_height)) x1, y1, x2, y2 = expand_bbox( x1, y1, x2, y2, big=expand_ratio, small=expand_ratio, frame_width=ori_width, frame_height=ori_height ) frame_mask[y1:y2, x1:x2] = 1 # 给每只手单独mask if not frame_valid: if for_vae: valid_flags[(i + 3) // 4] = False continue # 合并frame_mask到大mask if for_vae: compressed_idx = (i + 3) // 4 mask[compressed_idx] = torch.maximum(mask[compressed_idx], frame_mask) else: mask[i] = frame_mask if for_vae: for t, valid in enumerate(valid_flags): if not valid: mask[t] = 0 # 四帧内只要有一帧不合格，就清空对应压缩帧mask return mask def expand_bbox(x1, y1, x2, y2, big=0.1, small=0.4,max_expand_pixels=None, frame_width=None, frame_height=None): width = x2 - x1 height = y2 - y1 if width > height: width_expand_ratio = big height_expand_ratio = small else: width_expand_ratio = small height_expand_ratio = big expand_width = width * width_expand_ratio expand_height = height * height_expand_ratio if max_expand_pixels is not None: expand_width = min(expand_width, max_expand_pixels) expand_height = min(expand_height, max_expand_pixels) x1_new = max(0, x1 - expand_width) y1_new = max(0, y1 - expand_height) x2_new = x2 + expand_width y2_new = y2 + expand_height # 如果提供了图像宽度和高度，限制扩展后的坐标 if frame_width is not None: x2_new = min(x2_new, frame_width) if frame_height is not None: y2_new = min(y2_new, frame_height) return int(x1_new), int(y1_new), int(x2_new), int(y2_new) #ocr_bbox可能有多个, ocr_bbox需要expand,目标:ocr部分为0 def draw_ocr_mask(all_ocr_bbox, train_indices, real_t, ori_height, ori_width, for_vae = True): vae_t = 1 + (real_t-1) // 4 if len(all_ocr_bbox) == 0: if for_vae: return torch.ones((vae_t, ori_height, ori_width), dtype=torch.float32) else: return torch.ones((real_t, ori_height, ori_width), dtype=torch.float32) # 如果存在ocrbbox if for_vae: mask = torch.ones((vae_t, ori_height, ori_width), dtype=torch.float32) else: mask = torch.ones((real_t, ori_height, ori_width), dtype=torch.float32) for i, idx in enumerate(train_indices): bbox_list = all_ocr_bbox[idx] if len(bbox_list) == 0: continue for bbox in bbox_list: # 遍历所有的 bbox x1, y1, x2, y2 = map(int, bbox) x1, y1, x2, y2 = expand_bbox( x1, y1, x2, y2, max_expand_pixels=20, frame_width=ori_width, frame_height=ori_height ) if for_vae: compressed_idx = (i + 3) // 4 # 每 4 帧压缩成一帧 # 并集 mask[compressed_idx, y1:y2, x1:x2] = 0 else: mask[i, y1:y2, x1:x2] = 0 return mask def crop_tensors(tensors, bbox): if not bbox: return tensors x1, y1, x2, y2 = bbox return [t[..., y1:y2, x1:x2] for t in tensors] def get_bbox_from_mask(mask: torch.Tensor): if not mask.any(): return None combined_mask = mask.any(dim=0).squeeze(0) coords = combined_mask.nonzero(as_tuple=False) if coords.numel() == 0: return None y_min, x_min = coords.min(dim=0).values y_max, x_max = coords.max(dim=0).values return int(x_min), int(y_min), int(x_max), int(y_max) def apply_random_cropping_strategy( video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor, vid_face_bbox, ref_face_bbox, real_h, real_w, pose_cfg=0.2, refimg_cfg=0.0,refimg_crop=0.2, video_crop=0.2 ): p = random.random() if p < pose_cfg: # 场景A: CFG - 置零 pose 和 reference_frame pose = torch.zeros_like(pose) reference_frame = torch.zeros_like(reference_frame) reference_pose_tensor = torch.zeros_like(reference_pose_tensor) elif p < pose_cfg+refimg_cfg: # 场景B: CFG - 置零 reference_frame reference_frame = torch.zeros_like(reference_frame) reference_pose_tensor = torch.zeros_like(reference_pose_tensor) elif p < pose_cfg + refimg_cfg + video_crop and vid_face_bbox is not None: f_x1, f_y1, f_x2, f_y2 = vid_face_bbox H, W = video.shape[-2:] y1_crop = max(0, f_y1 - int((f_y2 - f_y1) * 0.1)) y2_crop = random.randint(int(H * 0.7), H) if y2_crop > y1_crop + 10: # 保证有效高度 crop_h = y2_crop - y1_crop crop_w = int(crop_h * (real_w / real_h)) x_start_range = max(0, f_x2 - crop_w) x_end_range = min(W - crop_w, f_x1) if x_start_range < x_end_range: x1_crop = random.randint(x_start_range, x_end_range) x2_crop = x1_crop + crop_w final_bbox = (x1_crop, y1_crop, x2_crop, y2_crop) video, pose, face_mask, ocr_mask, hand_mask = crop_tensors( [video, pose, face_mask, ocr_mask, hand_mask], final_bbox ) elif p < pose_cfg + refimg_cfg + video_crop + refimg_crop and ref_face_bbox is not None: f_x1, f_y1, f_x2, f_y2 = ref_face_bbox H, W = reference_frame.shape[-2:] y1_crop = max(0, f_y1 - int((f_y2 - f_y1) * 0.1)) y2_crop_min = f_y2 + int((f_y2 - f_y1) * 0.3) y2_crop = random.randint(min(y2_crop_min, H - 1), H - 1) if y2_crop > y1_crop + 10: crop_h = y2_crop - y1_crop crop_w = int(crop_h * (real_w / real_h)) x_start_range = max(0, f_x2 - crop_w) x_end_range = min(W - crop_w, f_x1) if x_start_range < x_end_range: x1_crop = random.randint(x_start_range, x_end_range) x2_crop = x1_crop + crop_w final_bbox = (x1_crop, y1_crop, x2_crop, y2_crop) reference_frame, reference_pose_tensor = crop_tensors([reference_frame, reference_pose_tensor], final_bbox) return video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor def generate_random_rectangle_mask(H, W, max_rects=1, max_size_ratio=0.2): mask = torch.zeros((1, 1, H, W), dtype=torch.uint8) for _ in range(random.randint(1, max_rects)): rect_h = random.randint(int(H * 0.05), int(H * max_size_ratio)) rect_w = random.randint(int(W * 0.05), int(W * max_size_ratio)) start_y = random.randint(0, H - rect_h) start_x = random.randint(0, W - rect_w) mask[:, :, start_y : start_y + rect_h, start_x : start_x + rect_w] = 1 return mask # use mask def apply_random_cropping_strategy_v2( video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor, vid_face_bbox, ref_face_bbox, real_h, real_w, pose_cfg=0.0, refimg_cfg=0.0, refimg_crop=0.0, random_zoomin=0.0,ref_keep=0.0, ): reference_frame_ori = reference_frame.clone() _, _, H, W = reference_frame.shape inpaint_mask = torch.zeros((1, 1, H, W),dtype=torch.uint8) p = random.random() if p < pose_cfg: pose = torch.zeros_like(pose) reference_frame = torch.zeros_like(reference_frame) reference_pose_tensor = torch.zeros_like(reference_pose_tensor) return video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor, inpaint_mask elif p < pose_cfg + refimg_cfg: reference_frame = torch.zeros_like(reference_frame) # reference_pose_tensor = torch.zeros_like(reference_pose_tensor) return video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor, inpaint_mask elif p < pose_cfg + refimg_cfg + refimg_crop and ref_face_bbox is not None: ratio_bottom = 0.9 f_x1, f_y1, f_x2, f_y2 = ref_face_bbox H, W = reference_frame.shape[-2:] y1_crop = max(0, f_y1 - int((f_y2 - f_y1) * 0.1)) y2_crop_min = f_y2 + int((f_y2 - f_y1) * 0.1) y2_crop_max = int(H * ratio_bottom) if y2_crop_max < y2_crop_min: # 极端情况下保证有合法区间 y2_crop_max = y2_crop_min y2_crop = random.randint(y2_crop_min, y2_crop_max) # y2_crop = random.randint(min(y2_crop_min, H - 1), H - 1) if y2_crop > y1_crop + 10: crop_h = y2_crop - y1_crop crop_w = int(crop_h * (real_w / real_h)) x_start_range = max(0, f_x2 - crop_w) x_end_range = min(W - crop_w, f_x1) if x_start_range < x_end_range: x1_crop = random.randint(x_start_range, x_end_range) x2_crop = x1_crop + crop_w final_bbox = (x1_crop, y1_crop, x2_crop, y2_crop) # x1, y1, x2, y2 = final_bbox mask_keep = torch.zeros((1, 1, H, W),dtype=torch.uint8) mask_keep[..., y1_crop:y2_crop, x1_crop:x2_crop] = 1 # 更新 inpaint_mask (bbox外为1，内为0) inpaint_mask = 1 - mask_keep # mask ref and ref pose reference_frame = reference_frame * mask_keep reference_pose_tensor = reference_pose_tensor * mask_keep else: if random.random() < 0.2: random_mask_to_add = generate_random_rectangle_mask(H, W).to(inpaint_mask.device) inpaint_mask = random_mask_to_add mask_keep = (1 - inpaint_mask).to(reference_frame.dtype) reference_frame = reference_frame * mask_keep reference_pose_tensor = reference_pose_tensor * mask_keep if random.random() < ref_keep: pose = torch.zeros_like(pose) reference_pose_tensor = torch.zeros_like(reference_pose_tensor) if video.shape[0] == 1: # T=1 video = reference_frame_ori return video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor, inpaint_mask if random_zoomin > 0.0: q = random.random() zoom_scale = 1.0 + random.uniform(0.0, 0.2) def zoom_in_center_crop(x, scale): Hx, Wx = x.shape[-2], x.shape[-1] newH = max(1, int(round(Hx * scale))) newW = max(1, int(round(Wx * scale))) if newH == Hx and newW == Wx: return x x_up = F.interpolate(x, size=(newH, newW), mode='nearest') top = max(0, (newH - Hx) // 2) left = max(0, (newW - Wx) // 2) return x_up[..., top:top + Hx, left:left + Wx] if q < random_zoomin * 0.5: reference_frame = zoom_in_center_crop(reference_frame, zoom_scale) reference_pose_tensor = zoom_in_center_crop(reference_pose_tensor, zoom_scale) inpaint_mask = zoom_in_center_crop(inpaint_mask.float(), zoom_scale).to(inpaint_mask.dtype) elif q < random_zoomin: video = zoom_in_center_crop(video, zoom_scale) pose = zoom_in_center_crop(pose, zoom_scale) face_mask = zoom_in_center_crop(face_mask.float(), zoom_scale).to(face_mask.dtype) ocr_mask = zoom_in_center_crop(ocr_mask.float(), zoom_scale).to(ocr_mask.dtype) hand_mask = zoom_in_center_crop(hand_mask.float(), zoom_scale).to(hand_mask.dtype) return video, pose, reference_frame, face_mask, ocr_mask, hand_mask, reference_pose_tensor, inpaint_mask def _calculate_center_crop_bbox(H, W, zoom_factor=0.7): crop_h = int(H * zoom_factor) crop_w = int(W * zoom_factor) y1 = (H - crop_h) // 2 x1 = (W - crop_w) // 2 return (x1, y1, x1 + crop_w, y1 + crop_h) def _calculate_smart_crop_bbox(bbox, H, W, real_h, real_w): ratio_bottom = 0.75 f_x1, f_y1, f_x2, f_y2 = bbox # 使用你之前为 refimg 设计的随机裁剪逻辑 y1_crop = max(0, f_y1 - int((f_y2 - f_y1) * 0.1)) y2_crop_min = f_y2 + int((f_y2 - f_y1) * 0.1) y2_crop_max = int(H * ratio_bottom) if y2_crop_max < y2_crop_min: # 极端情况下保证有合法区间 y2_crop_max = y2_crop_min y2_crop = random.randint(y2_crop_min, y2_crop_max) # y2_crop = random.randint(min(y2_crop_min, H - 1), H - 1) if y2_crop <= y1_crop + 10: return None crop_h = y2_crop - y1_crop crop_w = int(crop_h * (real_w / real_h)) x_start_range = max(0, f_x2 - crop_w) x_end_range = min(W - crop_w, f_x1) if x_start_range >= x_end_range: return None x1_crop = random.randint(x_start_range, x_end_range) x2_crop = x1_crop + crop_w return (x1_crop, y1_crop, x2_crop, y2_crop) # mask def apply_image_augmentation_strategy_v2( image, pose, reference_frame, reference_pose, tgt_face_mask, is_same_image, real_h, real_w, pose_cfg=0.0, refimg_cfg = 0.0, refimg_crop = 0.0, random_zoomin=0.0, ref_keep=0.0, ): reference_frame_ori = reference_frame.clone() _, _, H, W = reference_frame.shape inpaint_mask = torch.zeros((1, 1, H, W),dtype=torch.uint8) if random.random() < pose_cfg: pose = torch.zeros_like(pose) reference_frame = torch.zeros_like(reference_frame) reference_pose = torch.zeros_like(reference_pose) elif random.random() < pose_cfg + refimg_cfg: reference_frame = torch.zeros_like(reference_frame) # reference_pose = torch.zeros_like(reference_pose) # diff image elif refimg_crop == 0.0 or not is_same_image: return image, pose, reference_frame, reference_pose, tgt_face_mask, inpaint_mask else: bbox = get_bbox_from_mask(tgt_face_mask) final_bbox = None if bbox: final_bbox = _calculate_smart_crop_bbox(bbox, reference_frame.shape[-2], reference_frame.shape[-1], real_h, real_w) if final_bbox: x1_crop, y1_crop, x2_crop, y2_crop = final_bbox mask_keep = torch.zeros((1, 1, H, W),dtype=torch.uint8) mask_keep[..., y1_crop:y2_crop, x1_crop:x2_crop] = 1 inpaint_mask = 1 - mask_keep reference_frame = reference_frame * mask_keep reference_pose = reference_pose * mask_keep else: H, W = reference_frame.shape[-2:] center_crop_bbox = _calculate_center_crop_bbox(H, W) x1_crop, y1_crop, x2_crop, y2_crop = center_crop_bbox mask_keep = torch.zeros((1, 1, H, W),dtype=torch.uint8) mask_keep[..., y1_crop:y2_crop, x1_crop:x2_crop] = 1 inpaint_mask = 1 - mask_keep reference_frame = reference_frame * mask_keep reference_pose = reference_pose * mask_keep if random.random() < ref_keep: pose = torch.zeros_like(pose) reference_pose = torch.zeros_like(reference_pose) image = reference_frame_ori return image, pose, reference_frame, reference_pose, tgt_face_mask, inpaint_mask if random_zoomin > 0.0: q = random.random() zoom_scale = 1.0 + random.uniform(0.0, 0.3) def zoom_in_center_crop(x, scale): Hx, Wx = x.shape[-2], x.shape[-1] newH = max(1, int(round(Hx * scale))) newW = max(1, int(round(Wx * scale))) if newH == Hx and newW == Wx: return x x_up = F.interpolate(x, size=(newH, newW), mode='nearest') top = max(0, (newH - Hx) // 2) left = max(0, (newW - Wx) // 2) return x_up[..., top:top + Hx, left:left + Wx] if q < random_zoomin * 0.5: reference_frame = zoom_in_center_crop(reference_frame, zoom_scale) reference_frame = zoom_in_center_crop(reference_frame, zoom_scale) inpaint_mask = zoom_in_center_crop(inpaint_mask.float(), zoom_scale).to(inpaint_mask.dtype) elif q < random_zoomin: image = zoom_in_center_crop(image, zoom_scale) pose = zoom_in_center_crop(pose, zoom_scale) tgt_face_mask = zoom_in_center_crop(tgt_face_mask.float(), zoom_scale).to(tgt_face_mask.dtype) return image, pose, reference_frame, reference_pose, tgt_face_mask, inpaint_mask def apply_image_augmentation_strategy( image, pose, reference_frame, reference_pose, tgt_face_mask, is_same_image, real_h, real_w, pose_cfg=0.2, refimg_cfg = 0.0, refimg_crop = 0.0 ): if random.random() < pose_cfg: pose = torch.zeros_like(pose) reference_frame = torch.zeros_like(reference_frame) reference_pose = torch.zeros_like(reference_pose) return image, pose, reference_frame, reference_pose, tgt_face_mask if random.random() < pose_cfg + refimg_cfg: reference_frame = torch.zeros_like(reference_frame) reference_pose = torch.zeros_like(reference_pose) return image, pose, reference_frame, reference_pose, tgt_face_mask if refimg_crop == 0.0 or not is_same_image: return image, pose, reference_frame, reference_pose, tgt_face_mask bbox = get_bbox_from_mask(tgt_face_mask) final_bbox = None if bbox: final_bbox = _calculate_smart_crop_bbox(bbox, reference_frame.shape[-2], reference_frame.shape[-1], real_h, real_w) if final_bbox: if random.random() < 0.5: reference_frame, reference_pose = crop_tensors([reference_frame, reference_pose], final_bbox) else: image, pose, tgt_face_mask = crop_tensors([image, pose, tgt_face_mask], final_bbox) else: H, W = reference_frame.shape[-2:] center_crop_bbox = _calculate_center_crop_bbox(H, W) reference_frame, reference_pose = crop_tensors([reference_frame, reference_pose], center_crop_bbox) return image, pose, reference_frame, reference_pose, tgt_face_mask def crop_around_face(image, bbox, base_scale=0.3, jitter=2.0, aggressive_prob=0.6): H, W, _ = image.shape if len(bbox) == 0: return image x1, y1, x2, y2 = bbox face_cx = (x1 + x2) / 2 face_cy = (y1 + y2) / 2 face_top = min(y1, y2) face_bottom = max(y1, y2) face_h = face_bottom - face_top if random.random() > aggressive_prob: aspect_ratio = W / H crop_h = int(H * base_scale * random.uniform(1.0, jitter)) crop_w = int(crop_h * aspect_ratio) crop_h = min(crop_h, H) crop_w = min(crop_w, W) top = int(max(0, min(face_cy - crop_h * random.uniform(0.3, 0.5), H - crop_h))) left = int(max(0, min(face_cx - crop_w / 2, W - crop_w))) else: tight_scale = random.uniform(1.5, 2.5) crop_h = int(face_h * tight_scale) crop_w = int(crop_h * (W / H)) crop_h = min(crop_h, H) crop_w = min(crop_w, W) top = int(max(0, face_top - 0.15 * crop_h)) top = min(top, H - crop_h) left = int(max(0, face_cx - crop_w / 2)) left = min(left, W - crop_w) cropped = image[top:top + crop_h, left:left + crop_w, :] if cropped.shape[0] == 0 or cropped.shape[1] == 0: return image if cropped.shape[0] != crop_h or cropped.shape[1] != crop_w: return image return cropped # ========================== face region enhancement ======================== L_EYE_IDXS = list(range(36, 42)) R_EYE_IDXS = list(range(42, 48)) NOSE_TIP = 30 MOUTH_L = 48 MOUTH_R = 54 JAW_LINE = list(range(0, 17)) def _to_68x2(arr): arr = arr[:1] if arr.shape == (1, 68, 2): def to_orig(x): x = np.asarray(x, dtype=np.float64) if x.shape != (68, 2): raise ValueError("to_orig expects (68,2)") return x[np.newaxis, :, :] return arr[0].astype(np.float64), to_orig if arr.shape == (68, 2): def to_orig(x): x = np.asarray(x, dtype=np.float64) if x.shape != (68, 2): raise ValueError("to_orig expects (68,2)") return x return arr.astype(np.float64), to_orig if arr.shape == (2, 68): def to_orig(x): x = np.asarray(x, dtype=np.float64) if x.shape != (68, 2): raise ValueError("to_orig expects (68,2)") return x.T return arr.T.astype(np.float64), to_orig raise ValueError(f"faces shape {arr.shape} not supported; expected (1,68,2) or (68,2) or (2,68)") def _eye_center(face68, idxs): return face68[idxs].mean(axis=0) def _anchors(face68): le = _eye_center(face68, L_EYE_IDXS) re = _eye_center(face68, R_EYE_IDXS) nose = face68[NOSE_TIP] lm = face68[MOUTH_L] rm = face68[MOUTH_R] if re[0] < le[0]: le, re = re, le return np.stack([le, re, nose, lm, rm], axis=0) def _face_scale_only(src68, ref68, target_nose_pos, alpha=1.0, anchor_pairs=[[36, 45], [33, 8]],near_min=0.9, near_max=1.1, near_keep_prob=0.1, push_factor=1.2): src = np.asarray(src68, dtype=np.float64) ref = np.asarray(ref68, dtype=np.float64) center = _anchors(src).mean(axis=0) src_centered = src - center src_w = np.linalg.norm(src[anchor_pairs[0][0]] - src[anchor_pairs[0][1]]) ref_w = np.linalg.norm(ref[anchor_pairs[0][0]] - ref[anchor_pairs[0][1]]) src_h = np.linalg.norm(src[anchor_pairs[1][0]] - src[anchor_pairs[1][1]]) ref_h = np.linalg.norm(ref[anchor_pairs[1][0]] - ref[anchor_pairs[1][1]]) scale_x = ref_w / src_w if src_w > 1e-6 else 1.0 scale_y = ref_h / src_h if src_h > 1e-6 else 1.0 def push_if_near(val): if near_min < val < near_max: if np.random.rand() > near_keep_prob: if val >= 1.0: return max(near_max, val * push_factor) else: return min(near_min, val / push_factor) return val scale_x = push_if_near(scale_x) scale_y = push_if_near(scale_y) scaled_local = src_centered.copy() scaled_local[:, 0] *= (1 - alpha) + scale_x * alpha scaled_local[:, 1] *= (1 - alpha) + scale_y * alpha scaled_global = scaled_local + center nose_idx = NOSE_TIP current_nose = scaled_global[nose_idx] offset = target_nose_pos - current_nose scaled_global += offset return scaled_global def _face_roundify(src68, ref68, strength=1.0): src68 = np.asarray(src68, dtype=np.float64).copy() ref68 = np.asarray(ref68, dtype=np.float64) src_anchor = _anchors(src68) ref_anchor = _anchors(ref68) A = [] B = [] for sa, ra in zip(ref_anchor, src_anchor): # 注意顺序：ref -> src A.append([sa[0], sa[1], 1, 0, 0, 0]) A.append([0, 0, 0, sa[0], sa[1], 1]) B.append(ra[0]) B.append(ra[1]) A = np.array(A) B = np.array(B) x, *_ = np.linalg.lstsq(A, B, rcond=None) M_ref2src = np.array([ [x[0], x[1], x[2]], [x[3], x[4], x[5]] ]) ones = np.ones((ref68.shape[0], 1)) ref68_in_src = np.hstack([ref68, ones]) @ M_ref2src.T for idx in JAW_LINE: src68[idx] = (1-strength) * src68[idx] + strength * ref68_in_src[idx] return src68 def align_poses_by_limb_scaling_only( ref_pose, detected_poses, limb_change_prob=0.2, # 骨骼变动的概率（20%） face_change_prob=0.8, # 脸部变动的概率（80%） limb_alpha_range=(0.0, 0.8), # 骨骼缩放强度范围 face_alpha_range=(0.5, 1.0), # 脸部缩放强度范围 min_scale=0.5, max_scale=2.0, face_roundify_strength=1.0, seed=None): if not detected_poses: return [], False aligned_poses = copy.deepcopy(detected_poses) th = 1e-6 rng = np.random.RandomState(seed) if seed is not None else np.random ref_candidate = ref_pose['bodies']['candidate'] source_candidate_for_ratio = detected_poses[0]['bodies']['candidate'] def get_limb_length(points, p1_idx, p2_idx): return np.linalg.norm(points[p1_idx] - points[p2_idx]) def make_ratio(p1_idx, p2_idx, coef=1.0, alpha=0.0, near_min=0.9, near_max=1.1, near_keep_prob=0.1, push_factor=1.2): len_ref = get_limb_length(ref_candidate, p1_idx, p2_idx) len_src = get_limb_length(source_candidate_for_ratio, p1_idx, p2_idx) if len_src <= th or len_ref <= th: return 1.0 target = len_ref / len_src target = float(np.clip(target, min_scale, max_scale)) if near_min < target < near_max: if np.random.rand() > near_keep_prob: if target >= 1.0: target = max(near_max, target * push_factor) else: target = min(near_min, target / push_factor) return (1.0 * (1.0 - alpha) + target * alpha) * coef do_limb_change = (random.random() < limb_change_prob) do_face_change = (random.random() < face_change_prob) if do_limb_change: alpha_limb = rng.uniform(*limb_alpha_range) neck_ratio = make_ratio(0, 1, alpha=alpha_limb) r_eye_ratio = make_ratio(0,14, alpha=alpha_limb) l_eye_ratio = make_ratio(0,15, alpha=alpha_limb) r_ear_ratio = make_ratio(14,16, alpha=alpha_limb) l_ear_ratio = make_ratio(15,17, alpha=alpha_limb) else: neck_ratio = r_eye_ratio = l_eye_ratio = r_ear_ratio = l_ear_ratio = 1.0 has_ref_face = 'faces' in ref_pose and ref_pose['faces'] is not None if has_ref_face: ref68, _ = _to_68x2(ref_pose['faces']) for i, pose in enumerate(aligned_poses): candidate = pose['bodies']['candidate'] if do_limb_change: offset = (candidate[1] - candidate[0]) * (1.0 - neck_ratio) for idx in [0,14,15,16,17]: candidate[idx] += offset offset = (candidate[0] - candidate[14]) * (1.0 - r_eye_ratio) for idx in [14,16]: candidate[idx] += offset offset = (candidate[0] - candidate[15]) * (1.0 - l_eye_ratio) for idx in [15,17]: candidate[idx] += offset offset = (candidate[14] - candidate[16]) * (1.0 - r_ear_ratio) candidate[16] += offset offset = (candidate[15] - candidate[17]) * (1.0 - l_ear_ratio) candidate[17] += offset if has_ref_face and 'faces' in detected_poses[i] and detected_poses[i]['faces'] is not None: if do_face_change or do_limb_change: try: src68, to_orig = _to_68x2(detected_poses[i]['faces']) face_alpha = rng.uniform(*face_alpha_range) scaled68 = _face_scale_only(src68, ref68, candidate[0], alpha=face_alpha) scaled68 = _face_roundify(scaled68, ref68, strength=face_roundify_strength) pose['faces'] = to_orig(scaled68) except Exception as e: print('Face scaling error:', e) return aligned_poses, do_limb_change, do_face_change or do_limb_change def _euclid(p1, p2): return np.linalg.norm(p1 - p2) def align_pose_by_scale_and_translation(ref_pose: dict, tgt_pose: dict, conf_th: float = 0.5): aligned_pose = copy.deepcopy(tgt_pose) th = 1e-6 if 'bodies' not in ref_pose or 'bodies' not in tgt_pose: print("Warning: 'bodies' key not found in one of the poses. Returning original target pose.") return aligned_pose ref_kpt = ref_pose['bodies']['candidate'].astype(np.float32) tgt_kpt = aligned_pose['bodies']['candidate'].astype(np.float32) ref_sc = ref_pose['bodies'].get('score', np.ones(ref_kpt.shape[0])).astype(np.float32).reshape(-1) tgt_sc = tgt_pose['bodies'].get('score', np.ones(tgt_kpt.shape[0])).astype(np.float32).reshape(-1) ref_shoulder_valid = (ref_sc[2] >= conf_th) and (ref_sc[5] >= conf_th) tgt_shoulder_valid = (tgt_sc[2] >= conf_th) and (tgt_sc[5] >= conf_th) shoulder_ok = ref_shoulder_valid and tgt_shoulder_valid ref_hip_valid = (ref_sc[8] >= conf_th) and (ref_sc[11] >= conf_th) tgt_hip_valid = (tgt_sc[8] >= conf_th) and (tgt_sc[11] >= conf_th) hip_ok = ref_hip_valid and tgt_hip_valid ref_ear_valid = (ref_sc[16] >= conf_th) and (ref_sc[17] >= conf_th) tgt_ear_valid = (tgt_sc[16] >= conf_th) and (tgt_sc[17] >= conf_th) ear_ok = ref_ear_valid and tgt_ear_valid x_ratio, y_ratio = 1.0, 1.0 if shoulder_ok and hip_ok: # print("Scaling strategy: Using separate X (shoulder) and Y (torso) ratios.") ref_shoulder_w = abs(ref_kpt[5, 0] - ref_kpt[2, 0]) tgt_shoulder_w = abs(tgt_kpt[5, 0] - tgt_kpt[2, 0]) x_ratio = ref_shoulder_w / tgt_shoulder_w if tgt_shoulder_w > th else 1.0 ref_torso_h = abs(np.mean(ref_kpt[[8, 11], 1]) - np.mean(ref_kpt[[2, 5], 1])) tgt_torso_h = abs(np.mean(tgt_kpt[[8, 11], 1]) - np.mean(tgt_kpt[[2, 5], 1])) y_ratio = ref_torso_h / tgt_torso_h if tgt_torso_h > th else 1.0 elif shoulder_ok: # print("Scaling strategy: Fallback to unified ratio using shoulder distance.") ref_sh_dist = _euclid(ref_kpt[2], ref_kpt[5]) tgt_sh_dist = _euclid(tgt_kpt[2], tgt_kpt[5]) scale_ratio = ref_sh_dist / tgt_sh_dist if tgt_sh_dist > th else 1.0 x_ratio = y_ratio = scale_ratio else: # print("Scaling strategy: Fallback to unified ratio using ear distance.") ref_ear_dist = _euclid(ref_kpt[16], ref_kpt[17]) tgt_ear_dist = _euclid(tgt_kpt[16], tgt_kpt[17]) scale_ratio = ref_ear_dist / tgt_ear_dist if tgt_ear_dist > th else 1.0 x_ratio = y_ratio = scale_ratio # else: # print("Warning: No reliable keypoints for scaling. Using scale=1.0.") anchor_idx = 1 anchor_pt_before_scale = tgt_kpt[anchor_idx].copy() def scale(arr): if arr is not None and arr.size > 0: arr[..., 0] = anchor_pt_before_scale[0] + (arr[..., 0] - anchor_pt_before_scale[0]) * x_ratio arr[..., 1] = anchor_pt_before_scale[1] + (arr[..., 1] - anchor_pt_before_scale[1]) * y_ratio scale(tgt_kpt) scale(aligned_pose.get('faces')) scale(aligned_pose.get('hands')) offset = ref_kpt[anchor_idx] - tgt_kpt[anchor_idx] def translate(arr): if arr is not None and arr.size > 0: arr += offset translate(tgt_kpt) translate(aligned_pose.get('faces')) translate(aligned_pose.get('hands')) aligned_pose['bodies']['candidate'] = tgt_kpt return aligned_pose