Update app.py

app.py

@@ -1,54 +1,96 @@
+import cv2
+import numpy as np
+import math
+import os
 from flask import Flask, request, jsonify
-import cv2, numpy as np, math
 from cvzone.HandTrackingModule import HandDetector
 from cvzone.ClassificationModule import Classifier
 
-app = Flask(__name__)
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
 
-detector = HandDetector(maxHands=1)
-classifier = Classifier("Model_old/keras_model.h5", "Model_old/labels.txt")
+MODEL_PATH = "Model_old/keras_model.h5"
+LABELS_PATH = "Model_old/labels.txt"
 
+CONFIDENCE_THRESHOLD = 0.5
 IMG_SIZE = 300
 OFFSET = 20
+STABILITY_THRESHOLD = 3
+
+app = Flask(__name__)
+
+print("Loading hand detector...")
+detector = HandDetector(maxHands=1)
+
+print("Loading classifier...")
+classifier = Classifier(MODEL_PATH, LABELS_PATH)
+
+with open(LABELS_PATH) as f:
+    labels = [l.strip() for l in f.readlines()]
+
+last_gesture = None
+stable_gesture = None
+stable_count = 0
+
 
 @app.route("/predict", methods=["POST"])
 def predict():
+    global last_gesture, stable_gesture, stable_count
+
     if "frame" not in request.files:
         return jsonify({"gesture": None, "confidence": 0})
 
     file = request.files["frame"]
-    img = cv2.imdecode(np.frombuffer(file.read(), np.uint8), cv2.IMREAD_COLOR)
+    img_bytes = np.frombuffer(file.read(), np.uint8)
+    frame = cv2.imdecode(img_bytes, cv2.IMREAD_COLOR)
 
-    hands, _ = detector.findHands(img)
+    frame = cv2.flip(frame, 1)
+    hands, _ = detector.findHands(frame)
 
     if not hands:
+        stable_count = 0
+        stable_gesture = None
         return jsonify({"gesture": None, "confidence": 0})
 
     hand = hands[0]
     x, y, w, h = hand["bbox"]
 
     imgWhite = np.ones((IMG_SIZE, IMG_SIZE, 3), np.uint8) * 255
-    imgCrop = img[max(0,y-OFFSET):y+h+OFFSET, max(0,x-OFFSET):x+w+OFFSET]
+    imgCrop = frame[max(0, y - OFFSET):y + h + OFFSET,
+                    max(0, x - OFFSET):x + w + OFFSET]
+
+    if imgCrop.size == 0:
+        return jsonify({"gesture": None, "confidence": 0})
 
     aspectRatio = h / w
+
     if aspectRatio > 1:
         k = IMG_SIZE / h
         wCal = math.ceil(k * w)
         imgResize = cv2.resize(imgCrop, (wCal, IMG_SIZE))
-        imgWhite[:, (IMG_SIZE-wCal)//2:(IMG_SIZE+wCal)//2] = imgResize
+        wGap = math.ceil((IMG_SIZE - wCal) / 2)
+        imgWhite[:, wGap:wCal + wGap] = imgResize
     else:
         k = IMG_SIZE / w
         hCal = math.ceil(k * h)
         imgResize = cv2.resize(imgCrop, (IMG_SIZE, hCal))
-        imgWhite[(IMG_SIZE-hCal)//2:(IMG_SIZE+hCal)//2, :] = imgResize
+        hGap = math.ceil((IMG_SIZE - hCal) / 2)
+        imgWhite[hGap:hCal + hGap, :] = imgResize
 
     prediction, index = classifier.getPrediction(imgWhite, draw=False)
+    confidence = float(prediction[index])
+    gesture = labels[index]
+
+    if gesture == stable_gesture:
+        stable_count += 1
+    else:
+        stable_gesture = gesture
+        stable_count = 1
 
-    return jsonify({
-        "gesture": classifier.labels[index],
-        "confidence": float(prediction[index])
-    })
+    if stable_count >= STABILITY_THRESHOLD and confidence >= CONFIDENCE_THRESHOLD:
+        last_gesture = gesture
+        return jsonify({
+            "gesture": gesture,
+            "confidence": confidence
+        })
 
-# REQUIRED for Hugging Face Spaces
-if __name__ == "__main__":
-    app.run(host="0.0.0.0", port=7860)
+    return jsonify({"gesture": None, "confidence": 0})