Added PoseNet integration for faster processing

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "joints_detectors/posenet"]
+	path = joints_detectors/posenet
+	url = https://github.com/StanfordHCI/posenet-pytorch.git

tools/posenet_realtime.py

@@ -0,0 +1,53 @@
+import os
+import cv2
+from tqdm import tqdm
+import numpy as np
+from argparse import ArgumentParser
+import ipdb;pdb=ipdb.set_trace
+import time
+
+from tools.utils_posenet import getPosenetModel, getKptsFromImage
+poseModel = getPosenetModel()
+interface2D = getKptsFromImage
+from tools.utils import videopose_model_load as Model3Dload
+model3D = Model3Dload()
+from tools.utils import interface as VideoPoseInterface
+interface3D = VideoPoseInterface
+from tools.utils import draw_3Dimg, draw_2Dimg, videoInfo, resize_img
+
+def main(VideoName):
+    cap, cap_length = videoInfo(VideoName)
+    kpt2Ds = []
+    for i in tqdm(range(cap_length)):
+        _, frame = cap.read()
+        frame, W, H = resize_img(frame)
+
+        try:
+            t0 = time.time()
+            joint2D = interface2D(poseModel, frame)
+            print('PoseNet comsume {:0.3f} s'.format(time.time() - t0))
+        except Exception as e:
+            print(e)
+            continue
+
+        if i == 0:
+            for _ in range(30):
+                kpt2Ds.append(joint2D)
+        elif i < 30:
+            kpt2Ds.append(joint2D)
+            kpt2Ds.pop(0)
+        else:
+            kpt2Ds.append(joint2D)
+
+        joint3D = interface3D(model3D, np.array(kpt2Ds), W, H)
+        joint3D_item = joint3D[-1] #(17, 3)
+        draw_3Dimg(joint3D_item, frame, display=1, kpt2D=joint2D)
+
+if __name__ == '__main__':
+    parser = ArgumentParser()
+    parser.add_argument("-video", "--video_input", help="input video file name", default="/home/xyliu/Videos/sports/dance.mp4")
+    args = parser.parse_args()
+    VideoName = args.video_input
+    print('Input Video Name is ', VideoName)
+    main(VideoName)
+

tools/utils_posenet.py

@@ -0,0 +1,38 @@
+import os
+import sys
+
+import torch
+import numpy as np
+path = os.path.split(os.path.realpath(__file__))[0]
+posenet_path = os.path.join(path, '../joints_detectors/posenet/')
+sys.path.insert(0, posenet_path)
+import posenet
+
+def getPosenetModel():
+    model = posenet.load_model(101)
+    model = model.cuda()
+    return model
+
+def getKptsFromImage(model, input_image):
+    output_stride = model.output_stride
+    with torch.no_grad():
+        input_image, _, _ = posenet.utils._process_input(input_image, 1 / 3.0, output_stride)
+        input_image = torch.Tensor(input_image)
+        input_image = input_image.cuda()
+
+        heatmaps_result, offsets_result, displacement_fwd_result, displacement_bwd_result = model(input_image)
+
+        pose_scores, keypoint_scores, keypoint_coords = posenet.decode_multiple_poses(
+            heatmaps_result.squeeze(0),
+            offsets_result.squeeze(0),
+            displacement_fwd_result.squeeze(0),
+            displacement_bwd_result.squeeze(0),
+            output_stride=output_stride,
+            max_pose_detections=10,
+            min_pose_score=0.15)
+
+    #keypoint_coords *= output_scale
+    joint2 = np.zeros(keypoint_coords[0].shape)
+    joint2[:,0] = keypoint_coords[0][:,1].copy()
+    joint2[:,1] = keypoint_coords[0][:,0].copy()
+    return joint2