Create demo/vis.py

demo/vis.py

@@ -0,0 +1,356 @@
+import sys
+import argparse
+import cv2
+from lib.preprocess import h36m_coco_format, revise_kpts
+from lib.hrnet.gen_kpts import gen_video_kpts as hrnet_pose
+import os 
+import numpy as np
+import torch
+import torch.nn as nn
+import glob
+from tqdm import tqdm
+import copy
+
+sys.path.append(os.getcwd())
+from common.model_poseformer import PoseTransformerV2 as Model
+from common.camera import *
+
+import matplotlib
+import matplotlib.pyplot as plt 
+from mpl_toolkits.mplot3d import Axes3D
+import matplotlib.gridspec as gridspec
+
+plt.switch_backend('agg')
+matplotlib.rcParams['pdf.fonttype'] = 42
+matplotlib.rcParams['ps.fonttype'] = 42
+
+def show2Dpose(kps, img):
+    connections = [[0, 1], [1, 2], [2, 3], [0, 4], [4, 5],
+                   [5, 6], [0, 7], [7, 8], [8, 9], [9, 10],
+                   [8, 11], [11, 12], [12, 13], [8, 14], [14, 15], [15, 16]]
+
+    LR = np.array([0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0], dtype=bool)
+
+    lcolor = (255, 0, 0)
+    rcolor = (0, 0, 255)
+    thickness = 3
+
+    for j,c in enumerate(connections):
+        start = map(int, kps[c[0]])
+        end = map(int, kps[c[1]])
+        start = list(start)
+        end = list(end)
+        cv2.line(img, (start[0], start[1]), (end[0], end[1]), lcolor if LR[j] else rcolor, thickness)
+        cv2.circle(img, (start[0], start[1]), thickness=-1, color=(0, 255, 0), radius=3)
+        cv2.circle(img, (end[0], end[1]), thickness=-1, color=(0, 255, 0), radius=3)
+
+    return img
+
+
+def show3Dpose(vals, ax):
+    ax.view_init(elev=15., azim=70)
+
+    lcolor=(0,0,1)
+    rcolor=(1,0,0)
+
+    I = np.array( [0, 0, 1, 4, 2, 5, 0, 7,  8,  8, 14, 15, 11, 12, 8,  9])
+    J = np.array( [1, 4, 2, 5, 3, 6, 7, 8, 14, 11, 15, 16, 12, 13, 9, 10])
+
+    LR = np.array([0, 1, 0, 1, 0, 1, 0, 0, 0,   1,  0,  0,  1,  1, 0, 0], dtype=bool)
+
+    for i in np.arange( len(I) ):
+        x, y, z = [np.array( [vals[I[i], j], vals[J[i], j]] ) for j in range(3)]
+        ax.plot(x, y, z, lw=2, color = lcolor if LR[i] else rcolor)
+
+    RADIUS = 0.72
+    RADIUS_Z = 0.7
+
+    xroot, yroot, zroot = vals[0,0], vals[0,1], vals[0,2]
+    ax.set_xlim3d([-RADIUS+xroot, RADIUS+xroot])
+    ax.set_ylim3d([-RADIUS+yroot, RADIUS+yroot])
+    ax.set_zlim3d([-RADIUS_Z+zroot, RADIUS_Z+zroot])
+    ax.set_aspect('auto') # works fine in matplotlib==2.2.2
+
+    white = (1.0, 1.0, 1.0, 0.0)
+    ax.xaxis.set_pane_color(white) 
+    ax.yaxis.set_pane_color(white)
+    ax.zaxis.set_pane_color(white)
+
+    ax.tick_params('x', labelbottom = False)
+    ax.tick_params('y', labelleft = False)
+    ax.tick_params('z', labelleft = False)
+
+
+def get_pose2D(video_path, output_dir):
+    cap = cv2.VideoCapture(video_path)
+    width = cap.get(cv2.CAP_PROP_FRAME_WIDTH)
+    height = cap.get(cv2.CAP_PROP_FRAME_HEIGHT)
+
+    print('\nGenerating 2D pose...')
+    keypoints, scores = hrnet_pose(video_path, det_dim=416, num_peroson=1, gen_output=True)
+    keypoints, scores, valid_frames = h36m_coco_format(keypoints, scores)
+    re_kpts = revise_kpts(keypoints, scores, valid_frames)
+    print('Generating 2D pose successful!')
+
+    output_dir += 'input_2D/'
+    os.makedirs(output_dir, exist_ok=True)
+
+    output_npz = output_dir + 'keypoints.npz'
+    np.savez_compressed(output_npz, reconstruction=keypoints)
+
+
+def img2video(video_path, output_dir):
+    cap = cv2.VideoCapture(video_path)
+    fps = int(cap.get(cv2.CAP_PROP_FPS)) + 5
+
+    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+
+    names = sorted(glob.glob(os.path.join(output_dir + 'pose/', '*.png')))
+    img = cv2.imread(names[0])
+    size = (img.shape[1], img.shape[0])
+
+    videoWrite = cv2.VideoWriter(output_dir + video_name + '.mp4', fourcc, fps, size) 
+
+    for name in names:
+        img = cv2.imread(name)
+        videoWrite.write(img)
+
+    videoWrite.release()
+
+
+def showimage(ax, img):
+    ax.set_xticks([])
+    ax.set_yticks([]) 
+    plt.axis('off')
+    ax.imshow(img)
+
+
+def get_pose3D(video_path, output_dir):
+    args, _ = argparse.ArgumentParser().parse_known_args()
+    args.embed_dim_ratio, args.depth, args.frames = 32, 4, 243
+    args.number_of_kept_frames, args.number_of_kept_coeffs = 27, 27
+    args.pad = (args.frames - 1) // 2
+    args.previous_dir = 'checkpoint/'
+    args.n_joints, args.out_joints = 17, 17
+
+    ## Reload 
+    cuda_available = torch.cuda.is_available()
+    print(f"CUDA available in get_pose3D: {cuda_available}")
+    if cuda_available:
+        print(f"CUDA device count: {torch.cuda.device_count()}")
+        print(f"CUDA device name: {torch.cuda.get_device_name(0)}")
+    
+    device = torch.device('cuda' if cuda_available else 'cpu')
+    print(f"Using device: {device}")
+    
+    base_model = Model(args=args)
+    
+    # Always use DataParallel when CUDA is available (checkpoint expects it)
+    if cuda_available:
+        model = nn.DataParallel(base_model).to(device)
+    else:
+        model = base_model.to(device)
+
+    model_dict = model.state_dict()
+    # Put the pretrained model of PoseFormerV2 in 'checkpoint/']
+    # model_path = sorted(glob.glob(os.path.join(args.previous_dir, '27_243_45.2.bin')))
+    model_path = "./demo/lib/checkpoint/27_243_45.2.bin"
+
+    map_location = device
+    pre_dict = torch.load(model_path, map_location=map_location, weights_only=False)
+    
+    # Handle DataParallel checkpoint mismatch
+    state_dict = pre_dict['model_pos']
+    from collections import OrderedDict
+    new_state_dict = OrderedDict()
+    
+    # Check if we need to add or remove "module." prefix
+    checkpoint_has_module = any(k.startswith('module.') for k in state_dict.keys())
+    model_has_module = isinstance(model, nn.DataParallel)
+    
+    if checkpoint_has_module and not model_has_module:
+        # Remove "module." prefix
+        for k, v in state_dict.items():
+            name = k[7:] if k.startswith('module.') else k
+            new_state_dict[name] = v
+    elif not checkpoint_has_module and model_has_module:
+        # Add "module." prefix
+        for k, v in state_dict.items():
+            name = 'module.' + k if not k.startswith('module.') else k
+            new_state_dict[name] = v
+    else:
+        # No change needed
+        new_state_dict = state_dict
+    
+    model.load_state_dict(new_state_dict, strict=True)
+
+    model.eval()
+
+    ## input
+    keypoints = np.load(output_dir + 'input_2D/keypoints.npz', allow_pickle=True)['reconstruction']
+
+    cap = cv2.VideoCapture(video_path)
+    video_length = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+    ## 3D
+    print('\nGenerating 3D pose...')
+    keypoints_3D = []
+    for i in tqdm(range(video_length)):
+        ret, img = cap.read()
+        if img is None:
+            continue
+        img_size = img.shape
+
+        ## input frames
+        start = max(0, i - args.pad)
+        end =  min(i + args.pad, len(keypoints[0])-1)
+
+        input_2D_no = keypoints[0][start:end+1]
+        
+        left_pad, right_pad = 0, 0
+        if input_2D_no.shape[0] != args.frames:
+            if i < args.pad:
+                left_pad = args.pad - i
+            if i > len(keypoints[0]) - args.pad - 1:
+                right_pad = i + args.pad - (len(keypoints[0]) - 1)
+
+            input_2D_no = np.pad(input_2D_no, ((left_pad, right_pad), (0, 0), (0, 0)), 'edge')
+        
+        joints_left =  [4, 5, 6, 11, 12, 13]
+        joints_right = [1, 2, 3, 14, 15, 16]
+
+        # input_2D_no += np.random.normal(loc=0.0, scale=5, size=input_2D_no.shape)
+        input_2D = normalize_screen_coordinates(input_2D_no, w=img_size[1], h=img_size[0])  
+
+        input_2D_aug = copy.deepcopy(input_2D)
+        input_2D_aug[ :, :, 0] *= -1
+        input_2D_aug[ :, joints_left + joints_right] = input_2D_aug[ :, joints_right + joints_left]
+        input_2D = np.concatenate((np.expand_dims(input_2D, axis=0), np.expand_dims(input_2D_aug, axis=0)), 0)
+        # (2, 243, 17, 2)
+        
+        input_2D = input_2D[np.newaxis, :, :, :, :]
+
+        input_2D = torch.from_numpy(input_2D.astype('float32')).to(device)
+
+        N = input_2D.size(0)
+
+        ## estimation
+        output_3D_non_flip = model(input_2D[:, 0]) 
+        output_3D_flip     = model(input_2D[:, 1])
+        # [1, 1, 17, 3]
+
+        output_3D_flip[:, :, :, 0] *= -1
+        output_3D_flip[:, :, joints_left + joints_right, :] = output_3D_flip[:, :, joints_right + joints_left, :] 
+
+        output_3D = (output_3D_non_flip + output_3D_flip) / 2
+
+        output_3D[:, :, 0, :] = 0
+        post_out = output_3D[0, 0].cpu().detach().numpy()
+        keypoints_3D.append(post_out)
+        # print(f'Output 3D shape: {output_3D.shape}, post_out shape: {post_out.shape}, output 3D sample: {output_3D[0]}, post out sample: {post_out}')
+
+        rot =  [0.1407056450843811, -0.1500701755285263, -0.755240797996521, 0.6223280429840088]
+        rot = np.array(rot, dtype='float32')
+        post_out = camera_to_world(post_out, R=rot, t=0)
+        post_out[:, 2] -= np.min(post_out[:, 2])
+
+        input_2D_no = input_2D_no[args.pad]
+
+        ## 2D
+        image = show2Dpose(input_2D_no, copy.deepcopy(img))
+
+        output_dir_2D = output_dir +'pose2D/'
+        os.makedirs(output_dir_2D, exist_ok=True)
+        cv2.imwrite(output_dir_2D + str(('%04d'% i)) + '_2D.png', image)
+
+        ## 3D
+        fig = plt.figure(figsize=(9.6, 5.4))
+        gs = gridspec.GridSpec(1, 1)
+        gs.update(wspace=-0.00, hspace=0.05) 
+        ax = plt.subplot(gs[0], projection='3d')
+        show3Dpose( post_out, ax)
+
+        output_dir_3D = output_dir +'pose3D/'
+        os.makedirs(output_dir_3D, exist_ok=True)
+        plt.savefig(output_dir_3D + str(('%04d'% i)) + '_3D.png', dpi=200, format='png', bbox_inches = 'tight')
+        plt.clf()
+        plt.close(fig)
+    
+    output_npz = output_dir + 'keypoints_3D.npz'
+    np.savez_compressed(output_npz, reconstruction=keypoints_3D)
+    print('Generating 3D pose successful!')
+
+    ## all
+    image_dir = 'results/' 
+    image_2d_dir = sorted(glob.glob(os.path.join(output_dir_2D, '*.png')))
+    image_3d_dir = sorted(glob.glob(os.path.join(output_dir_3D, '*.png')))
+
+    print('\nGenerating demo...')
+    for i in tqdm(range(len(image_2d_dir))):
+        image_2d = plt.imread(image_2d_dir[i])
+        image_3d = plt.imread(image_3d_dir[i])
+
+        ## crop
+        edge = (image_2d.shape[1] - image_2d.shape[0]) // 2
+        image_2d = image_2d[:, edge:image_2d.shape[1] - edge]
+
+        edge = 130
+        image_3d = image_3d[edge:image_3d.shape[0] - edge, edge:image_3d.shape[1] - edge]
+
+        ## show
+        font_size = 12
+        fig = plt.figure(figsize=(15.0, 5.4))
+        ax = plt.subplot(121)
+        showimage(ax, image_2d)
+        ax.set_title("Input", fontsize = font_size)
+
+        ax = plt.subplot(122)
+        showimage(ax, image_3d)
+        ax.set_title("Reconstruction", fontsize = font_size)
+
+        ## save
+        output_dir_pose = output_dir +'pose/'
+        os.makedirs(output_dir_pose, exist_ok=True)
+        plt.subplots_adjust(top=1, bottom=0, right=1, left=0, hspace=0, wspace=0)
+        plt.margins(0, 0)
+        plt.savefig(output_dir_pose + str(('%04d'% i)) + '_pose.png', dpi=200, bbox_inches = 'tight')
+        plt.clf()
+        plt.close(fig)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--video', type=str, default='sample_video.mp4', help='input video')
+    parser.add_argument('--gpu', type=str, default='0', help='GPU device ID (set CUDA_VISIBLE_DEVICES before running if needed)')
+    args = parser.parse_args()
+
+    # Note: CUDA_VISIBLE_DEVICES must be set BEFORE importing torch
+    # Since torch is imported at the top, setting it here won't work
+    # Set it in your environment before running: $env:CUDA_VISIBLE_DEVICES="0" (PowerShell) or export CUDA_VISIBLE_DEVICES=0 (bash)
+    
+    # Verify CUDA availability
+    print(f"CUDA available: {torch.cuda.is_available()}")
+    if torch.cuda.is_available():
+        print(f"CUDA device count: {torch.cuda.device_count()}")
+        print(f"Current device: {torch.cuda.current_device()}")
+        print(f"Device name: {torch.cuda.get_device_name(0)}")
+        if "CUDA_VISIBLE_DEVICES" in os.environ:
+            print(f"CUDA_VISIBLE_DEVICES={os.environ['CUDA_VISIBLE_DEVICES']}")
+    else:
+        print("WARNING: CUDA is not available!")
+        print("This might be because:")
+        print("  1. CUDA_VISIBLE_DEVICES was set incorrectly")
+        print("  2. PyTorch was installed without CUDA support")
+        print("  3. GPU drivers are not installed")
+        print("\nTo use GPU, set CUDA_VISIBLE_DEVICES BEFORE running Python:")
+        print("  PowerShell: $env:CUDA_VISIBLE_DEVICES='0'")
+        print("  Bash: export CUDA_VISIBLE_DEVICES=0")
+        print("\nOr don't set it at all to use the default GPU")
+
+    video_path = './demo/video/' + args.video
+    video_name = video_path.split('/')[-1].split('.')[0]
+    output_dir = './demo/output/' + video_name + '/'
+
+    get_pose2D(video_path, output_dir)
+    get_pose3D(video_path, output_dir)
+    img2video(video_path, output_dir)
+    print('Generating demo successful!')