How to Add Confidence Score for Each Attribute on a Custom Person Attribute Recognition Model Using Intel® DL Streamer?

Users are experiencing difficulties in adding confidence scores for each attribute in a custom person attribute recognition model using Intel® DL Streamer.

1. Create attribute_grouper.py script with the code below:

   import numpy as np
   # Full attribute list in EXACT model output order (verify with benchmark_app)
   ATTRIBUTES_LIST = [
     "accessoryHat",    # 0
     "hairLong",      # 1
     "hairShort",     # 2
     "upperBodyShortSleeve", # 3
     "upperBodyBlack",   # 4
     "upperBodyBlue",   # 5
     "upperBodyBrown",   # 6
     "upperBodyGreen",   # 7
     "upperBodyGrey",   # 8
     "upperBodyOrange",  # 9
     "upperBodyPink",   # 10
     "upperBodyPurple",  # 11
     "upperBodyRed",    # 12
     "upperBodyWhite",   # 13
     "upperBodyYellow",  # 14
     "upperBodyLongSleeve", # 15
     "lowerBodyShorts",  # 16
     "lowerBodyShortSkirt", # 17
     "lowerBodyBlack",   # 18
     "lowerBodyBlue",   # 19
     "lowerBodyBrown",   # 20
     "lowerBodyGreen",   # 21
     "lowerBodyGrey",   # 22
     "lowerBodyOrange",  # 23
     "lowerBodyPink",   # 24
     "lowerBodyPurple",  # 25
     "lowerBodyRed",    # 26
     "lowerBodyWhite",   # 27
     "lowerBodyYellow",  # 28
     "lowerBodyLongSkirt", # 29
     "footwearLeatherShoes", # 30
     "footwearSandals",  # 31
     "footwearShoes",   # 32
     "footwearSneaker",  # 33
     "carryingBackpack",  # 34
     "carryingMessengerBag", # 35
     "carryingLuggageCase", # 36
     "carryingSuitcase",  # 37
     "personalLess30",   # 38
     "personalLess45",   # 39
     "personalLess60",   # 40
     "personalLarger60",  # 41
     "personalLess15",   # 42
     "personalMale",    # 43
     "personalFemale"   # 44
   ]
   # Configure your 9 attribute groups here
   ATTRIBUTE_GROUPS = {
     # Group 1: Head Attributes
     "Head": {
       "attributes": ["accessoryHat", "hairLong", "hairShort"],
       "display_template": "Head: {attribute}",
       "threshold": 0.4 # Lower threshold for head items
     },
     # Group 2: Upper Body Color
     "UpperColor": {
       "attributes": ["upperBodyBlack", "upperBodyBlue", "upperBodyBrown",
              "upperBodyGreen", "upperBodyGrey", "upperBodyOrange",
              "upperBodyPink", "upperBodyPurple", "upperBodyRed",
              "upperBodyWhite", "upperBodyYellow"],
       "display_template": "Upper: {attribute}",
       "threshold": 0.5
     },
     # Group 3: Upper Body Style
     "UpperStyle": {
       "attributes": ["upperBodyShortSleeve", "upperBodyLongSleeve"],
       "display_template": "Sleeve: {attribute}",
       "threshold": 0.5,
       "mutually_exclusive": True # Only one can be true
     },
     # Group 4: Lower Body Color
     "LowerColor": {
       "attributes": ["lowerBodyBlack", "lowerBodyBlue", "lowerBodyBrown",
              "lowerBodyGreen", "lowerBodyGrey", "lowerBodyOrange",
              "lowerBodyPink", "lowerBodyPurple", "lowerBodyRed",
              "lowerBodyWhite", "lowerBodyYellow"],
       "display_template": "Lower: {attribute}",
       "threshold": 0.5
     },
     # Group 5: Lower Body Style
     "LowerStyle": {
       "attributes": ["lowerBodyShorts", "lowerBodyShortSkirt", "lowerBodyLongSkirt"],
       "display_template": "Bottom: {attribute}",
       "threshold": 0.5,
       "mutually_exclusive": True
     },
     # Group 6: Footwear
     "Footwear": {
       "attributes": ["footwearLeatherShoes", "footwearSandals", 
              "footwearShoes", "footwearSneaker"],
       "display_template": "Shoes: {attribute}",
       "threshold": 0.5
     },
     # Group 7: Carrying Items
     "Carrying": {
       "attributes": ["carryingBackpack", "carryingMessengerBag",
              "carryingLuggageCase", "carryingSuitcase"],
       "display_template": "Bag: {attribute}",
       "threshold": 0.4 # Lower threshold for rare items
     },
     # Group 8: Age Group
     "Age": {
       "attributes": ["personalLess15", "personalLess30", 
              "personalLess45", "personalLess60", "personalLarger60"],
       "display_template": "Age: {attribute}",
       "threshold": 0.5,
       "mutually_exclusive": True
     },
     # Group 9: Gender
     "Gender": {
       "attributes": ["personalMale", "personalFemale"],
       "display_template": "Gender: {attribute}",
       "threshold": 0.7, # Higher threshold for gender
       "mutually_exclusive": True
     }
   }
   def process_frame(frame):
     for roi in frame.regions():
       # 1. Extract raw predictions
       attr_predictions = {}
       for tensor in roi.tensors():
         if tensor.format() == "ANY":
           data = tensor.data()
           if len(data) == len(ATTRIBUTES_LIST):
             for i, confidence in enumerate(data):
               attr_predictions[ATTRIBUTES_LIST[i]] = float(confidence)
       # 2. Process each group
       results = []
       for group_name, config in ATTRIBUTE_GROUPS.items():
         best_attr = None
         max_conf = 0
         for attribute in config["attributes"]:
           conf = attr_predictions.get(attribute, 0)
           # Skip if mutually exclusive and we already have a high-confidence result
           if config.get("mutually_exclusive", False) and best_attr and max_conf > config["threshold"]:
             continue
           if conf > max_conf and conf > config["threshold"]:
             max_conf = conf
             best_attr = attribute
         if best_attr:
           display_text = config["display_template"].format(attribute=best_attr)
           results.append((display_text, max_conf))
       # 3. Update ROI labels (shows all groups with detected attributes)
       if results:
         # Combine all results into one label for cleaner display
         combined_label = " | ".join([text for text, _ in results])
         roi.set_label(combined_label, min([conf for _, conf in results]))
     return True
   # For debugging (enable in development)
   DEBUG_MODE = False
   if DEBUG_MODE:
     def debug_print_predictions(predictions):
       print("\nRaw Predictions:")
       for attr, conf in predictions.items():
         if conf > 0.3: # Only show attributes with >30% confidence
           print(f"{attr}: {conf:.2f}")
     # Add this inside process_frame after getting attr_predictions
     debug_print_predictions(attr_predictions)

2. Run the following GStreamer pipeline in Terminal:

   gst-launch-1.0 \
   filesrc location=video.mp4 ! decodebin ! videoconvert ! \
   gvadetect model=person_vehicle.xml device=CPU ! queue ! \
   gvainference model=attribute_model.xml device=CPU inference-region=roi ! queue ! \
   gvapython module=attribute_grouper.py ! \
   gvawatermark ! videoconvert ! autovideosink sync=false