pose_align.py

import numpy as np
import torch
import copy
import cv2
import os
import moviepy.video.io.ImageSequenceClip
from datetime import datetime
import gc

from pose.script.dwpose import DWposeDetector, draw_pose
from pose.script.util import size_calculate, warpAffine_kps
from downloading_weights import download_models

# ZeroGPU
import spaces


'''
    Detect dwpose from img, then align it by scale parameters
    img: frame from the pose video
    detector: DWpose
    scales: scale parameters
'''
class PoseAlignmentInference:
    def __init__(self,
                 model_dir,
                 output_dir):
        self.detector = None
        self.model_paths = {
            "det_ckpt": os.path.join(model_dir, "dwpose", "yolox_l_8x8_300e_coco.pth"),
            "pose_ckpt": os.path.join(model_dir, "dwpose", "dw-ll_ucoco_384.pth")
        }
        self.config_paths = {
            "pose_config": os.path.join("pose", "config", "dwpose-l_384x288.py"),
            "det_config": os.path.join("pose", "config", "yolox_l_8xb8-300e_coco.py"),
        }
        self.model_dir = model_dir
        self.output_dir = os.path.join(output_dir, "pose_alignment")
        if not os.path.exists(self.output_dir):
            os.makedirs(self.output_dir)

    @spaces.GPU(duration=120)
    def align_pose(
        self,
        vidfn: str,
        imgfn_refer: str,
        detect_resolution: int,
        image_resolution: int,
        align_frame: int,
        max_frame: int,
    ):
        download_models(model_dir=self.model_dir)
        output_filename = "pose_temp"
        outfn=os.path.abspath(os.path.join(self.output_dir, f'{output_filename}_demo.mp4'))
        outfn_align_pose_video=os.path.abspath(os.path.join(self.output_dir, f'{output_filename}.mp4'))

        video = cv2.VideoCapture(vidfn)
        width= video.get(cv2.CAP_PROP_FRAME_WIDTH)
        height= video.get(cv2.CAP_PROP_FRAME_HEIGHT)

        total_frame= video.get(cv2.CAP_PROP_FRAME_COUNT)
        fps= video.get(cv2.CAP_PROP_FPS)

        print("height:", height)
        print("width:", width)
        print("fps:", fps)

        H_in, W_in  = height, width
        H_out, W_out = size_calculate(H_in,W_in, detect_resolution)
        H_out, W_out = size_calculate(H_out,W_out, image_resolution)

        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        self.detector = DWposeDetector(
            det_config = self.config_paths["det_config"],
            det_ckpt = self.model_paths["det_ckpt"],
            pose_config = self.config_paths["pose_config"],
            pose_ckpt = self.model_paths["pose_ckpt"],
            keypoints_only=False
        )
        detector = self.detector.to(device)

        refer_img = cv2.imread(imgfn_refer)
        output_refer, pose_refer = detector(refer_img,detect_resolution=detect_resolution, image_resolution=image_resolution, output_type='cv2',return_pose_dict=True)
        body_ref_img  = pose_refer['bodies']['candidate']
        hands_ref_img = pose_refer['hands']
        faces_ref_img = pose_refer['faces']
        output_refer = cv2.cvtColor(output_refer, cv2.COLOR_RGB2BGR)


        skip_frames = align_frame
        max_frame = max_frame
        pose_list, video_frame_buffer, video_pose_buffer = [], [], []


        cap = cv2.VideoCapture('2.mp4')     # 读取视频
        while cap.isOpened():               # 当视频被打开时：
            ret, frame = cap.read()         # 读取视频，读取到的某一帧存储到frame，若是读取成功，ret为True，反之为False
            if ret:                         # 若是读取成功
                cv2.imshow('frame', frame)  # 显示读取到的这一帧画面
                key = cv2.waitKey(25)       # 等待一段时间，并且检测键盘输入
                if key == ord('q'):         # 若是键盘输入'q',则退出，释放视频
                    cap.release()           # 释放视频
                    break
            else:
                cap.release()
        cv2.destroyAllWindows()             # 关闭所有窗口


        for i in range(max_frame):
            ret, img = video.read()
            if img is None:
                break
            else:
                if i < skip_frames:
                    continue
                video_frame_buffer.append(img)

            # estimate scale parameters by the 1st frame in the video
            if i==skip_frames:
                output_1st_img, pose_1st_img = detector(img, detect_resolution, image_resolution, output_type='cv2', return_pose_dict=True)
                body_1st_img  = pose_1st_img['bodies']['candidate']
                hands_1st_img = pose_1st_img['hands']
                faces_1st_img = pose_1st_img['faces']

                '''
                计算逻辑:
                1. 先把 ref 和 pose 的高 resize 到一样，且都保持原来的长宽比。
                2. 用点在图中的实际坐标来计算。
                3. 实际计算中，把h的坐标归一化到 [0, 1],  w为[0, W/H]
                4. 由于 dwpose 的输出本来就是归一化的坐标，所以h不需要变，w要乘W/H
                注意：dwpose 输出是 (w, h)
                '''

                # h不变，w缩放到原比例
                ref_H, ref_W = refer_img.shape[0], refer_img.shape[1]
                ref_ratio = ref_W / ref_H
                body_ref_img[:, 0]  = body_ref_img[:, 0] * ref_ratio
                hands_ref_img[:, :, 0] = hands_ref_img[:, :, 0] * ref_ratio
                faces_ref_img[:, :, 0] = faces_ref_img[:, :, 0] * ref_ratio

                video_ratio = width / height
                body_1st_img[:, 0]  = body_1st_img[:, 0] * video_ratio
                hands_1st_img[:, :, 0] = hands_1st_img[:, :, 0] * video_ratio
                faces_1st_img[:, :, 0] = faces_1st_img[:, :, 0] * video_ratio

                # scale
                align_args = dict()

                dist_1st_img = np.linalg.norm(body_1st_img[0]-body_1st_img[1])   # 0.078
                dist_ref_img = np.linalg.norm(body_ref_img[0]-body_ref_img[1])   # 0.106
                align_args["scale_neck"] = dist_ref_img / dist_1st_img  # align / pose = ref / 1st

                dist_1st_img = np.linalg.norm(body_1st_img[16]-body_1st_img[17])
                dist_ref_img = np.linalg.norm(body_ref_img[16]-body_ref_img[17])
                align_args["scale_face"] = dist_ref_img / dist_1st_img

                dist_1st_img = np.linalg.norm(body_1st_img[2]-body_1st_img[5])  # 0.112
                dist_ref_img = np.linalg.norm(body_ref_img[2]-body_ref_img[5])  # 0.174
                align_args["scale_shoulder"] = dist_ref_img / dist_1st_img

                dist_1st_img = np.linalg.norm(body_1st_img[2]-body_1st_img[3])  # 0.895
                dist_ref_img = np.linalg.norm(body_ref_img[2]-body_ref_img[3])  # 0.134
                s1 = dist_ref_img / dist_1st_img
                dist_1st_img = np.linalg.norm(body_1st_img[5]-body_1st_img[6])
                dist_ref_img = np.linalg.norm(body_ref_img[5]-body_ref_img[6])
                s2 = dist_ref_img / dist_1st_img
                align_args["scale_arm_upper"] = (s1+s2)/2 # 1.548

                dist_1st_img = np.linalg.norm(body_1st_img[3]-body_1st_img[4])
                dist_ref_img = np.linalg.norm(body_ref_img[3]-body_ref_img[4])
                s1 = dist_ref_img / dist_1st_img
                dist_1st_img = np.linalg.norm(body_1st_img[6]-body_1st_img[7])
                dist_ref_img = np.linalg.norm(body_ref_img[6]-body_ref_img[7])
                s2 = dist_ref_img / dist_1st_img
                align_args["scale_arm_lower"] = (s1+s2)/2

                # hand
                dist_1st_img = np.zeros(10)
                dist_ref_img = np.zeros(10)

                dist_1st_img[0] = np.linalg.norm(hands_1st_img[0,0]-hands_1st_img[0,1])
                dist_1st_img[1] = np.linalg.norm(hands_1st_img[0,0]-hands_1st_img[0,5])
                dist_1st_img[2] = np.linalg.norm(hands_1st_img[0,0]-hands_1st_img[0,9])
                dist_1st_img[3] = np.linalg.norm(hands_1st_img[0,0]-hands_1st_img[0,13])
                dist_1st_img[4] = np.linalg.norm(hands_1st_img[0,0]-hands_1st_img[0,17])
                dist_1st_img[5] = np.linalg.norm(hands_1st_img[1,0]-hands_1st_img[1,1])
                dist_1st_img[6] = np.linalg.norm(hands_1st_img[1,0]-hands_1st_img[1,5])
                dist_1st_img[7] = np.linalg.norm(hands_1st_img[1,0]-hands_1st_img[1,9])
                dist_1st_img[8] = np.linalg.norm(hands_1st_img[1,0]-hands_1st_img[1,13])
                dist_1st_img[9] = np.linalg.norm(hands_1st_img[1,0]-hands_1st_img[1,17])

                dist_ref_img[0] = np.linalg.norm(hands_ref_img[0,0]-hands_ref_img[0,1])
                dist_ref_img[1] = np.linalg.norm(hands_ref_img[0,0]-hands_ref_img[0,5])
                dist_ref_img[2] = np.linalg.norm(hands_ref_img[0,0]-hands_ref_img[0,9])
                dist_ref_img[3] = np.linalg.norm(hands_ref_img[0,0]-hands_ref_img[0,13])
                dist_ref_img[4] = np.linalg.norm(hands_ref_img[0,0]-hands_ref_img[0,17])
                dist_ref_img[5] = np.linalg.norm(hands_ref_img[1,0]-hands_ref_img[1,1])
                dist_ref_img[6] = np.linalg.norm(hands_ref_img[1,0]-hands_ref_img[1,5])
                dist_ref_img[7] = np.linalg.norm(hands_ref_img[1,0]-hands_ref_img[1,9])
                dist_ref_img[8] = np.linalg.norm(hands_ref_img[1,0]-hands_ref_img[1,13])
                dist_ref_img[9] = np.linalg.norm(hands_ref_img[1,0]-hands_ref_img[1,17])

                ratio = 0
                count = 0
                for i in range (10):
                    if dist_1st_img[i] != 0:
                        ratio = ratio + dist_ref_img[i]/dist_1st_img[i]
                        count = count + 1
                if count!=0:
                    align_args["scale_hand"] = (ratio/count+align_args["scale_arm_upper"]+align_args["scale_arm_lower"])/3
                else:
                    align_args["scale_hand"] = (align_args["scale_arm_upper"]+align_args["scale_arm_lower"])/2

                # body
                dist_1st_img = np.linalg.norm(body_1st_img[1] - (body_1st_img[8] + body_1st_img[11])/2 )
                dist_ref_img = np.linalg.norm(body_ref_img[1] - (body_ref_img[8] + body_ref_img[11])/2 )
                align_args["scale_body_len"]=dist_ref_img / dist_1st_img

                dist_1st_img = np.linalg.norm(body_1st_img[8]-body_1st_img[9])
                dist_ref_img = np.linalg.norm(body_ref_img[8]-body_ref_img[9])
                s1 = dist_ref_img / dist_1st_img
                dist_1st_img = np.linalg.norm(body_1st_img[11]-body_1st_img[12])
                dist_ref_img = np.linalg.norm(body_ref_img[11]-body_ref_img[12])
                s2 = dist_ref_img / dist_1st_img
                align_args["scale_leg_upper"] = (s1+s2)/2

                dist_1st_img = np.linalg.norm(body_1st_img[9]-body_1st_img[10])
                dist_ref_img = np.linalg.norm(body_ref_img[9]-body_ref_img[10])
                s1 = dist_ref_img / dist_1st_img
                dist_1st_img = np.linalg.norm(body_1st_img[12]-body_1st_img[13])
                dist_ref_img = np.linalg.norm(body_ref_img[12]-body_ref_img[13])
                s2 = dist_ref_img / dist_1st_img
                align_args["scale_leg_lower"] = (s1+s2)/2

                ####################
                ####################
                # need adjust nan
                for k,v in align_args.items():
                    if np.isnan(v):
                        align_args[k]=1

                # centre offset (the offset of key point 1)
                offset = body_ref_img[1] - body_1st_img[1]


            # pose align
            pose_img, pose_ori = detector(img, detect_resolution, image_resolution, output_type='cv2', return_pose_dict=True)
            video_pose_buffer.append(pose_img)
            pose_align = self.align_img(img, pose_ori, align_args, detect_resolution, image_resolution)


            # add centre offset
            pose = pose_align
            pose['bodies']['candidate'] = pose['bodies']['candidate'] + offset
            pose['hands'] = pose['hands'] + offset
            pose['faces'] = pose['faces'] + offset


            # h不变，w从绝对坐标缩放回0-1 注意这里要回到ref的坐标系
            pose['bodies']['candidate'][:, 0] = pose['bodies']['candidate'][:, 0] / ref_ratio
            pose['hands'][:, :, 0] = pose['hands'][:, :, 0] / ref_ratio
            pose['faces'][:, :, 0] = pose['faces'][:, :, 0] / ref_ratio
            pose_list.append(pose)

        # stack
        body_list  = [pose['bodies']['candidate'][:18] for pose in pose_list]
        body_list_subset = [pose['bodies']['subset'][:1] for pose in pose_list]
        hands_list = [pose['hands'][:2] for pose in pose_list]
        faces_list = [pose['faces'][:1] for pose in pose_list]

        body_seq         = np.stack(body_list       , axis=0)
        body_seq_subset  = np.stack(body_list_subset, axis=0)
        hands_seq        = np.stack(hands_list      , axis=0)
        faces_seq        = np.stack(faces_list      , axis=0)


        # concatenate and paint results
        H = 768 # paint height
        W1 = int((H/ref_H * ref_W)//2 *2)
        W2 = int((H/height * width)//2 *2)
        result_demo = [] # = Writer(args, None, H, 3*W1+2*W2, outfn, fps)
        result_pose_only = [] # Writer(args, None, H, W1, args.outfn_align_pose_video, fps)
        for i in range(len(body_seq)):
            pose_t={}
            pose_t["bodies"]={}
            pose_t["bodies"]["candidate"]=body_seq[i]
            pose_t["bodies"]["subset"]=body_seq_subset[i]
            pose_t["hands"]=hands_seq[i]
            pose_t["faces"]=faces_seq[i]

            ref_img = cv2.cvtColor(refer_img, cv2.COLOR_RGB2BGR)
            ref_img = cv2.resize(ref_img, (W1, H))
            ref_pose= cv2.resize(output_refer, (W1, H))

            output_transformed = draw_pose(
                pose_t,
                int(H_in*1024/W_in),
                1024,
                draw_face=False,
                )
            output_transformed = cv2.cvtColor(output_transformed, cv2.COLOR_BGR2RGB)
            output_transformed = cv2.resize(output_transformed, (W1, H))

            video_frame = cv2.resize(video_frame_buffer[i], (W2, H))
            video_pose  = cv2.resize(video_pose_buffer[i], (W2, H))

            res = np.concatenate([ref_img, ref_pose, output_transformed, video_frame, video_pose], axis=1)
            result_demo.append(res)
            result_pose_only.append(output_transformed)

        print(f"pose_list len: {len(pose_list)}")
        clip = moviepy.video.io.ImageSequenceClip.ImageSequenceClip(result_demo, fps=fps)
        clip.write_videofile(outfn, fps=fps)
        clip = moviepy.video.io.ImageSequenceClip.ImageSequenceClip(result_pose_only, fps=fps)
        clip.write_videofile(outfn_align_pose_video, fps=fps)
        print('pose align done')
        self.release_vram()
        return outfn_align_pose_video, outfn

    def release_vram(self):
        if self.detector is not None:
            del self.detector
            self.detector = None
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
            gc.collect()

    @staticmethod
    def align_img(img, pose_ori, scales, detect_resolution, image_resolution):

        body_pose = copy.deepcopy(pose_ori['bodies']['candidate'])
        hands = copy.deepcopy(pose_ori['hands'])
        faces = copy.deepcopy(pose_ori['faces'])

        '''
        计算逻辑:
        0. 该函数内进行绝对变换，始终保持人体中心点 body_pose[1] 不变
        1. 先把 ref 和 pose 的高 resize 到一样，且都保持原来的长宽比。
        2. 用点在图中的实际坐标来计算。
        3. 实际计算中，把h的坐标归一化到 [0, 1],  w为[0, W/H]
        4. 由于 dwpose 的输出本来就是归一化的坐标，所以h不需要变，w要乘W/H
        注意：dwpose 输出是 (w, h)
        '''

        # h不变，w缩放到原比例
        H_in, W_in, C_in = img.shape
        video_ratio = W_in / H_in
        body_pose[:, 0] = body_pose[:, 0] * video_ratio
        hands[:, :, 0] = hands[:, :, 0] * video_ratio
        faces[:, :, 0] = faces[:, :, 0] * video_ratio

        # scales of 10 body parts
        scale_neck = scales["scale_neck"]
        scale_face = scales["scale_face"]
        scale_shoulder = scales["scale_shoulder"]
        scale_arm_upper = scales["scale_arm_upper"]
        scale_arm_lower = scales["scale_arm_lower"]
        scale_hand = scales["scale_hand"]
        scale_body_len = scales["scale_body_len"]
        scale_leg_upper = scales["scale_leg_upper"]
        scale_leg_lower = scales["scale_leg_lower"]

        scale_sum = 0
        count = 0
        scale_list = [scale_neck, scale_face, scale_shoulder, scale_arm_upper, scale_arm_lower, scale_hand,
                      scale_body_len, scale_leg_upper, scale_leg_lower]
        for i in range(len(scale_list)):
            if not np.isinf(scale_list[i]):
                scale_sum = scale_sum + scale_list[i]
                count = count + 1
        for i in range(len(scale_list)):
            if np.isinf(scale_list[i]):
                scale_list[i] = scale_sum / count

        # offsets of each part
        offset = dict()
        offset["14_15_16_17_to_0"] = body_pose[[14, 15, 16, 17], :] - body_pose[[0], :]
        offset["3_to_2"] = body_pose[[3], :] - body_pose[[2], :]
        offset["4_to_3"] = body_pose[[4], :] - body_pose[[3], :]
        offset["6_to_5"] = body_pose[[6], :] - body_pose[[5], :]
        offset["7_to_6"] = body_pose[[7], :] - body_pose[[6], :]
        offset["9_to_8"] = body_pose[[9], :] - body_pose[[8], :]
        offset["10_to_9"] = body_pose[[10], :] - body_pose[[9], :]
        offset["12_to_11"] = body_pose[[12], :] - body_pose[[11], :]
        offset["13_to_12"] = body_pose[[13], :] - body_pose[[12], :]
        offset["hand_left_to_4"] = hands[1, :, :] - body_pose[[4], :]
        offset["hand_right_to_7"] = hands[0, :, :] - body_pose[[7], :]

        # neck
        c_ = body_pose[1]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_neck)

        neck = body_pose[[0], :]
        neck = warpAffine_kps(neck, M)
        body_pose[[0], :] = neck

        # body_pose_up_shoulder
        c_ = body_pose[0]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_face)

        body_pose_up_shoulder = offset["14_15_16_17_to_0"] + body_pose[[0], :]
        body_pose_up_shoulder = warpAffine_kps(body_pose_up_shoulder, M)
        body_pose[[14, 15, 16, 17], :] = body_pose_up_shoulder

        # shoulder
        c_ = body_pose[1]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_shoulder)

        body_pose_shoulder = body_pose[[2, 5], :]
        body_pose_shoulder = warpAffine_kps(body_pose_shoulder, M)
        body_pose[[2, 5], :] = body_pose_shoulder

        # arm upper left
        c_ = body_pose[2]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_arm_upper)

        elbow = offset["3_to_2"] + body_pose[[2], :]
        elbow = warpAffine_kps(elbow, M)
        body_pose[[3], :] = elbow

        # arm lower left
        c_ = body_pose[3]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_arm_lower)

        wrist = offset["4_to_3"] + body_pose[[3], :]
        wrist = warpAffine_kps(wrist, M)
        body_pose[[4], :] = wrist

        # hand left
        c_ = body_pose[4]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_hand)

        hand = offset["hand_left_to_4"] + body_pose[[4], :]
        hand = warpAffine_kps(hand, M)
        hands[1, :, :] = hand

        # arm upper right
        c_ = body_pose[5]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_arm_upper)

        elbow = offset["6_to_5"] + body_pose[[5], :]
        elbow = warpAffine_kps(elbow, M)
        body_pose[[6], :] = elbow

        # arm lower right
        c_ = body_pose[6]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_arm_lower)

        wrist = offset["7_to_6"] + body_pose[[6], :]
        wrist = warpAffine_kps(wrist, M)
        body_pose[[7], :] = wrist

        # hand right
        c_ = body_pose[7]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_hand)

        hand = offset["hand_right_to_7"] + body_pose[[7], :]
        hand = warpAffine_kps(hand, M)
        hands[0, :, :] = hand

        # body len
        c_ = body_pose[1]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_body_len)

        body_len = body_pose[[8, 11], :]
        body_len = warpAffine_kps(body_len, M)
        body_pose[[8, 11], :] = body_len

        # leg upper left
        c_ = body_pose[8]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_leg_upper)

        knee = offset["9_to_8"] + body_pose[[8], :]
        knee = warpAffine_kps(knee, M)
        body_pose[[9], :] = knee

        # leg lower left
        c_ = body_pose[9]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_leg_lower)

        ankle = offset["10_to_9"] + body_pose[[9], :]
        ankle = warpAffine_kps(ankle, M)
        body_pose[[10], :] = ankle

        # leg upper right
        c_ = body_pose[11]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_leg_upper)

        knee = offset["12_to_11"] + body_pose[[11], :]
        knee = warpAffine_kps(knee, M)
        body_pose[[12], :] = knee

        # leg lower right
        c_ = body_pose[12]
        cx = c_[0]
        cy = c_[1]
        M = cv2.getRotationMatrix2D((cx, cy), 0, scale_leg_lower)

        ankle = offset["13_to_12"] + body_pose[[12], :]
        ankle = warpAffine_kps(ankle, M)
        body_pose[[13], :] = ankle

        # none part
        body_pose_none = pose_ori['bodies']['candidate'] == -1.
        hands_none = pose_ori['hands'] == -1.
        faces_none = pose_ori['faces'] == -1.

        body_pose[body_pose_none] = -1.
        hands[hands_none] = -1.
        nan = float('nan')
        if len(hands[np.isnan(hands)]) > 0:
            print('nan')
        faces[faces_none] = -1.

        # last check nan -> -1.
        body_pose = np.nan_to_num(body_pose, nan=-1.)
        hands = np.nan_to_num(hands, nan=-1.)
        faces = np.nan_to_num(faces, nan=-1.)

        # return
        pose_align = copy.deepcopy(pose_ori)
        pose_align['bodies']['candidate'] = body_pose
        pose_align['hands'] = hands
        pose_align['faces'] = faces

        return pose_align