Dev

evaluate.sh

@@ -4,9 +4,9 @@
 # pose_number is how many number of keypoints we want to see in a person
 # pose_threshold is the score we set to filter the keypoints whose score is small 
 # ground_truth='/export/home/cyh/mygithub/PoseGCN/data/annotations/val_2017/'
-ground_truth='${PGPT_ROOT}/data/demodataset/annotations/'
-predictions='${PGPT_ROOT}/results/demo'
-results='test'
+ground_truth='/PGPT/data/demodataset/annotations'
+predictions='/PGPT/results/demo'
+results='/PGPT/results/evaluate'
 pose_number=0
 pose_threshold=0.5
 

inference/config.py → inference/config_old.py

@@ -1,31 +1,33 @@
 
 class Config():
-    root = '${PGPT_ROOT}'
+    root = '/PGPT'
+
     # save_dir is the loaction where we store the results
     save_dir = root + '/results/demo'
+
     # json_path_detection is the loaction where we store the detection results
     json_path_detection = root + '/results/demo_detection.json'
-
-    # gt_json_path is the gound truth of the validiation, all the ground_truth are in one file
+    
+    # gt_json_path is the ground truth of the validiation, all the ground_truth are in one file
     gt_json_path = root + '/data/demo_val.json'
-
+    
     # the data folder of the PoseTrack dataset
     data_folder = root + '/data/demodataset'
-
+    
     # the path of the location where we store the video
-    video_path = root 
-
+    video_path = save_dir
+    
     # the path of the track model
     track_model = root + '/models/tracker.pth'
-
+    
     # the path of the pose estimation model
     pose_model = root + '/models/pose_gcn.pth.tar'
-
+    
     # pose config file location
     pose_cfg = root + '/cfgs/pose_res152.yaml'
-
+    
     # the path of the embedding model
     embedding_model = root + '/models/embedding_model.pth'
-
+    
     def __init__(self):
         print('Using the config class at', __file__)

inference/inference.py

@@ -9,7 +9,13 @@
 
 from tqdm import tqdm
 import random
-from config import Config
+
+#from config import Config
+import sys
+sys.path.append('../lib')
+from config import cfg
+from config import update_config
+
 from track_and_detect_new import Track_And_Detect
 
 '''
@@ -51,13 +57,21 @@
 }
 '''
 match_list=[13,12,14,9,8,10,7,11,6,3,2,4,1,5,0]
-config = Config()
+#config = Config()
 def parseArgs():
 	parser = argparse.ArgumentParser(description="Evaluation of Pose Estimation and Tracking (PoseTrack)")
+	parser.add_argument('--cfg', type=str, required=True) #added by alnguyen
 	parser.add_argument("-t", "--detection_thresh",dest = 'det_thresh',required=False, default=0.4, type= float)
 	parser.add_argument("-p", "--pos_thresh",dest = 'pose_thresh',required=False, default=0, type= float)
 	parser.add_argument("-v", "--vis_flag",dest = 'vis_flag',required=False, default=False, type= bool)
-	return parser.parse_args()
+	parser.add_argument('opts', 
+                        help='Modify config options using the command-line',
+                        default=None,
+                        nargs=argparse.REMAINDER) #added by alnguyen
+
+	args = parser.parse_args()
+
+	return args
 
 class DateEncoder(json.JSONEncoder):
 	def default(self, obj):
@@ -67,18 +81,22 @@ def default(self, obj):
 			return obj.tolist()
 		return json.JSONEncoder.default(self, obj)
 
-def track_test(args, gpu_id=0):
+def track_test():
 
+	args = parseArgs()
 	pose_vis_thresh = args.pose_thresh
 	detection_score_thresh = args.det_thresh
 	vis_flag = args.vis_flag
-	json_path = config.json_path_detection
+
+	update_config(cfg, args)
+	gpu_id = cfg.GPU_ID
+	json_path = cfg.INPUT.JSON_DETECTION_PATH
 	# Change temporially 
-	save_dir = config.save_dir
+	save_dir = cfg.OUTPUT.SAVE_DIR
 
-	gt_json_path = config.gt_json_path
-	data_folder = config.data_folder
-	video_path = config.video_path
+	gt_json_path = cfg.INPUT.GT_JSON_PATH
+	data_folder = cfg.INPUT.DATA_FOLDER
+	video_path = cfg.OUTPUT.VIDEO_PATH
 
 	print('----------------------------------------------------------------------------------')
 	print('Detection_score_thresh: {}    Vis_flag: {}'.format(detection_score_thresh, vis_flag))
@@ -87,24 +105,27 @@ def track_test(args, gpu_id=0):
 	if not os.path.exists(save_dir):
 		os.makedirs(save_dir)
 
-	# Load the Detection Results 
+	# Load the Detection Results (demo_detection.json)
 	with open(json_path,'r') as f:
 		bbox_dict = json.load(f)
 
 	# Create the Tracker
-	tracker = Track_And_Detect(gpu_id=gpu_id, track_model=config.track_model, pose_model=config.pose_model, embedding_model=config.embedding_model)
+	track_model=cfg.INPUT.TRACK_MODEL
+	pose_model=cfg.INPUT.POSE_MODEL
+	embedding_model=cfg.INPUT.EMBEDDING_MODEL
+	tracker = Track_And_Detect(gpu_id=gpu_id, track_model=track_model, pose_model=pose_model, embedding_model=embedding_model)
 
-	# Load the Ground Truth to get the right video keys
+	# Load the Ground Truth to get the right video keys (demo_val.json)
 	with open(gt_json_path,'r') as f:
 		gt_dict = json.load(f)
 
 
 	video_keys = gt_dict.keys()
 	pbar = tqdm(range(len(video_keys)))
-	for video_name in video_keys:
+	for video_name in video_keys: #in demo_val.json
 		pbar.update(1)
 		frame_dict = bbox_dict[video_name]
-		video_name = video_name.replace('.json','')
+		#video_name = video_name.replace('.json','')
 		video_json = {'annolist':[]}
 		save_path = os.path.join(save_dir, video_name+'.json')
 		idx =0
@@ -124,7 +145,7 @@ def track_test(args, gpu_id=0):
 					if not os.path.exists(video_path):
 						os.makedirs(video_path)
 					video_store_name = video_path + '/{}.mp4'
-					videoWriter = cv2.VideoWriter(video_store_name.format(video_name),fourcc,10,(im_W,im_H))
+					videoWriter = cv2.VideoWriter(video_store_name.format(video_name+'-pgpt'),fourcc,10,(im_W,im_H))
 				final_list = tracker.init_tracker(frame,det_list)
 			else:	
 				track_list = tracker.multi_track(frame)
@@ -154,5 +175,4 @@ def track_test(args, gpu_id=0):
 
 
 if __name__ == "__main__":
-	args = parseArgs()
-	track_test(args=args)
+	track_test()

inference/pose_estimation_graph.py

@@ -56,7 +56,7 @@
 '''
 class PoseNet(object):
 	def __init__(self, gpu_id=0, model_path=None):
-		self.cfg_file='${PGPT_ROOT}/cfgs/pose_res152.yaml'
+		self.cfg_file='/PGPT/cfgs/pose_res152.yaml'
 		self.flag = 0
 
 		update_config(self.cfg_file)

inference/skeleton_visulize.py

@@ -5,12 +5,17 @@
 import cv2
 import numpy as np
 from cv2_color import Color
+import glob
+#from config import Config
+import sys
+sys.path.append('../lib')
+from config import cfg
+from config import update_config
 
 # The match list from the results to the test
 match_list=[13,12,14,9,8,10,7,11,6,3,2,4,1,5,0]
 color = Color(flag='bgr')
 
-
 def draw_limb(image, kps, color):
     def draw_line(head, tail):
         if head == [] or tail == []:
@@ -33,7 +38,17 @@ def draw_line(head, tail):
     for h, t in limbs:
         draw_line(kps[h], kps[t])
 
+def parseArgs():
+	parser = argparse.ArgumentParser(description="Visualizing the results")
+	parser.add_argument('--cfg', type=str, required=True) #added by alnguyen
+	parser.add_argument('opts', 
+                        help='Modify config options using the command-line',
+                        default=None,
+                        nargs=argparse.REMAINDER) #added by alnguyen
+
+	args = parser.parse_args()
 
+	return args
 
 def demo(image_dir, result_dir, save_dir):
     """
@@ -42,12 +57,14 @@ def demo(image_dir, result_dir, save_dir):
     save_dir: the loaction where we store the result videos
     """
 
-    json_files = os.listdir(result_dir)
+    #json_files = os.listdir(result_dir)
+    json_files = glob.glob(result_dir + "/*.json")
+    json_files = [json_file.split('/')[-1] for json_file in json_files]
 
     pbar = tqdm(range(len(json_files)))
     for json_name in json_files:
 
-        video_name = json_name.replace('.json','_new')
+        video_name = json_name.replace('.json','-pgpt')
 
         video_folder = os.path.join(save_dir, video_name)
 
@@ -58,11 +75,13 @@ def demo(image_dir, result_dir, save_dir):
             old_annolist = json.load(f)['annolist']
         pbar.set_description('Visulizing video {}'.format(video_name))
         color_list = color.get_random_color_list()
+        j = 0
         for i,annotation in enumerate(old_annolist):
             color_flag = 0
             frame_name = annotation['image'][0]['name']
-            frame_store_path = video_folder + '/{}'.format(frame_name.split('/')[-1])
-            frame_path = os.path.join(image_dir,frame_name)
+            #annotation['image_id'] #10010010103=frame_id
+            frame_store_path = os.path.join(video_folder, frame_name.split('/')[-1]) #file
+            frame_path = os.path.join(image_dir,frame_name) #file for read
             frame = cv2.imread(frame_path)
             im_H, im_W, im_C = frame.shape
             if i==0:
@@ -73,30 +92,37 @@ def demo(image_dir, result_dir, save_dir):
                 if len(anno['annopoints']) == 0:
                     continue
                 old_point_list = anno['annopoints'][0]['point']
-
                 xmin, xmax, ymin, ymax, track_id = anno['x1'][0], anno['x2'][0], anno['y1'][0], anno['y2'][0], anno['track_id'][0]
                 color_flag = int(track_id) % 16
-
                 kps = [[] for _ in range(15)]
                 for pose in old_point_list:
-
                     pose_id, pose_x, pose_y, = pose['id'][0], pose['x'][0], pose['y'][0] 
                     kps[pose_id] = (int(pose_x), int(pose_y))
-                    cv2.circle(frame,(int(pose_x),int(pose_y)), 3 ,color_list[color_flag], -1)  
+                    cv2.circle(frame,(int(pose_x),int(pose_y)), 3 ,color_list[color_flag], -1)
                 draw_limb(frame, kps, color_list[color_flag])
                 cv2.rectangle(frame, (int(xmin),int(ymin)), (int(xmax),int(ymax)), color_list[color_flag], 3)            
                 cv2.putText(frame, 'id:' + str(track_id), (int(xmin),int(ymin)), cv2.FONT_HERSHEY_SIMPLEX, 1, color_list[color_flag], 2)
             videoWriter.write(frame)
             cv2.imwrite(frame_store_path, frame)
+        #videoWriter.release()
         pbar.update(1)
     pbar.close()
 
 
 if __name__ == '__main__':
     print('Visualizing the results')
-    image_dir = '${PGPT_ROOT}/data/demodataset/'
-    result_dir = '${PGPT_ROOT}/results/demo/'
-    save_dir = '${PGPT_ROOT}/results/render/'
+    #parser.add_argument("--image_dir", type=str, default=cfg.INPUT.DATA_FOLDER) # /PGPT/data/demodataset
+    #parser.add_argument("--result_dir", type=str, default=cfg.OUTPUT.SAVE_DIR) # /PGPT/results/demo
+    #parser.add_argument("--save_dir", type=str, default=cfg.OUTPUT.VIDEO_PATH) # /PGPT/results/demo
+
+
+    args = parseArgs()
+    update_config(cfg, args)
+
+    image_dir = cfg.INPUT.DATA_FOLDER
+    result_dir = cfg.OUTPUT.SAVE_DIR
+    save_dir = cfg.OUTPUT.VIDEO_PATH
+
     if not os.path.exists(save_dir):
         os.makedirs(save_dir)
     demo(image_dir, result_dir, save_dir)

inference/track_and_detect_new.py

@@ -15,7 +15,7 @@
 from tracker import SiamFCTracker
 
 from match import Matcher
-from model.nms.nms_wrapper import nms
+#from model.nms.nms_wrapper import nms
 
 class Track_And_Detect(object):
 	effective_track_thresh = 0.5

lib/config/__init__.py

@@ -0,0 +1,9 @@
+# ------------------------------------------------------------------------------
+# Copyright (c) Microsoft
+# Licensed under the MIT License.
+# The code is based on HigherHRNet-Human-Pose-Estimation.
+# (https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation)
+# ------------------------------------------------------------------------------
+
+from .default import _C as cfg
+from .default import update_config

lib/config/default.py

@@ -0,0 +1,53 @@
+import os
+from yacs.config import CfgNode as CN
+
+
+_C = CN()
+
+_C.GPU_ID = 0 #ne marche pas pour les autres valeurs (1,2,3)
+_C.SYSTEM = CN()
+# Number of GPUS to use in the experiment
+_C.SYSTEM.NUM_GPUS = 4
+# Number of workers for doing things
+_C.SYSTEM.NUM_WORKERS = 4
+
+_C.SYSTEM.PGPT_ROOT = '/PGPT'
+# pose config file location
+_C.SYSTEM.POSE_CONFIG = '/PGPT/cfgs/pose_res152.yaml'
+
+_C.INPUT = CN()
+# json from detection and pose, images directory
+_C.INPUT.JSON_DETECTION_PATH = '/PGPT/results/demo_detection.json'
+# gt_json_path is the ground truth of the validiation, all the ground_truth are in one file
+_C.INPUT.GT_JSON_PATH = '/PGPT/data/demo_val.json'
+# the data folder of the PoseTrack dataset
+_C.INPUT.DATA_FOLDER =  '/PGPT/data/demodataset'
+# the path of the track model
+_C.INPUT.TRACK_MODEL = '/PGPT/models/tracker.pth'
+# the path of the pose estimation model
+_C.INPUT.POSE_MODEL = '/PGPT/models/pose_gcn.pth.tar'
+_C.INPUT.EMBEDDING_MODEL = '/PGPT/models/embedding_model.pth'
+
+_C.OUTPUT = CN()
+# where we store the results
+_C.OUTPUT.SAVE_DIR = '/PGPT/results/demo'
+# the path of the location where we store the video
+_C.OUTPUT.VIDEO_PATH = '/PGPT/results/demo/demo-pgpt.mp4'
+
+def get_cfg_defaults():
+  """Get a yacs CfgNode object with default values for my_project."""
+  # Return a clone so that the defaults will not be altered
+  # This is for the "local variable" use pattern
+  return _C.clone()
+
+def update_config(cfg, args):
+    cfg.defrost()
+    cfg.merge_from_file(args.cfg)
+    cfg.merge_from_list(args.opts)
+    cfg.freeze()
+
+
+if __name__ == '__main__':
+    import sys
+    with open(sys.argv[1], 'w') as f:
+        print(_C, file=f)

lib/poseval/py/eval_helpers.py

@@ -227,8 +227,9 @@ def process_arguments(argv):
   elif len(argv)<3 or len(argv)>4:
     help()
 
-  gt_file = argv[1]
-  pred_file = argv[2]
+  gt_dir = argv[1]
+  pred_dir = argv[2]
+  return gt_dir, pred_dir, mode
 
   if not os.path.exists(gt_file):
     help('Given ground truth directory does not exist!\n')
@@ -282,7 +283,6 @@ def load_data(argv):
 
   return gtFramesAll, prFramesAll
 
-
 def cleanupData(gtFramesAll,prFramesAll):
 
   # remove all GT frames with empty annorects and remove corresponding entries from predictions
@@ -377,6 +377,7 @@ def load_data_dir(argv):
   if not os.path.exists(pred_dir):
     help('Given prediction directory ' + pred_dir + ' does not exist!\n')
   filenames = glob.glob(gt_dir + "/*.json")
+
   gtFramesAll = []
   prFramesAll = []