Spaces:
Sleeping
Sleeping
Upload app.py
Browse files
app.py
CHANGED
@@ -9,11 +9,10 @@ import numpy as np
|
|
9 |
REALTIME_MODELS = {
|
10 |
"YOLOS Tiny (ultra-rapide)": "hustvl/yolos-tiny",
|
11 |
"DETR ResNet-50": "facebook/detr-resnet-50",
|
12 |
-
"YOLOS Small": "hustvl/yolos-small"
|
13 |
-
"Conditional DETR": "microsoft/conditional-detr-resnet-50"
|
14 |
}
|
15 |
|
16 |
-
# Variables globales
|
17 |
current_detector = None
|
18 |
current_model_name = None
|
19 |
|
@@ -36,26 +35,29 @@ def load_detector(model_name):
|
|
36 |
return current_detector
|
37 |
|
38 |
@spaces.GPU
|
39 |
-
def
|
40 |
-
"""
|
41 |
-
|
42 |
-
|
43 |
if image is None:
|
|
|
44 |
return None
|
45 |
|
46 |
try:
|
47 |
-
#
|
48 |
-
detector = load_detector(model_choice)
|
49 |
-
|
50 |
-
# Convertir en PIL Image si c'est un array numpy
|
51 |
if isinstance(image, np.ndarray):
|
52 |
pil_image = Image.fromarray(image)
|
53 |
else:
|
54 |
pil_image = image
|
|
|
|
|
55 |
|
56 |
-
#
|
|
|
|
|
|
|
|
|
57 |
original_size = pil_image.size
|
58 |
-
max_size = 480 # Taille réduite pour plus de vitesse
|
59 |
|
60 |
if max(original_size) > max_size:
|
61 |
ratio = max_size / max(original_size)
|
@@ -65,8 +67,11 @@ def detect_objects_live(image, model_choice, confidence_threshold):
|
|
65 |
resized_image = pil_image
|
66 |
ratio = 1.0
|
67 |
|
68 |
-
|
|
|
|
|
69 |
detections = detector(resized_image)
|
|
|
70 |
|
71 |
# Filtrer par confiance
|
72 |
filtered_detections = [
|
@@ -74,9 +79,11 @@ def detect_objects_live(image, model_choice, confidence_threshold):
|
|
74 |
if det['score'] >= confidence_threshold
|
75 |
]
|
76 |
|
77 |
-
print(f"
|
|
|
|
|
78 |
|
79 |
-
# Ajuster les coordonnées
|
80 |
for det in filtered_detections:
|
81 |
if ratio != 1.0:
|
82 |
det['box']['xmin'] = int(det['box']['xmin'] / ratio)
|
@@ -85,29 +92,28 @@ def detect_objects_live(image, model_choice, confidence_threshold):
|
|
85 |
det['box']['ymax'] = int(det['box']['ymax'] / ratio)
|
86 |
|
87 |
# Dessiner les détections
|
88 |
-
|
89 |
|
90 |
-
|
|
|
91 |
|
92 |
except Exception as e:
|
93 |
-
print(f"❌ Erreur: {e}")
|
|
|
|
|
94 |
return image
|
95 |
|
96 |
def draw_detections(image, detections):
|
97 |
-
"""Dessine les
|
98 |
-
if not detections:
|
99 |
-
return image
|
100 |
-
|
101 |
-
# Créer une copie pour dessiner
|
102 |
img_copy = image.copy()
|
103 |
draw = ImageDraw.Draw(img_copy)
|
104 |
|
105 |
-
# Couleurs
|
106 |
-
colors = ["#FF0000", "#00FF00", "#0000FF", "#FFFF00", "#FF00FF"
|
107 |
|
|
|
108 |
try:
|
109 |
-
|
110 |
-
font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 20)
|
111 |
except:
|
112 |
font = ImageFont.load_default()
|
113 |
|
@@ -116,76 +122,40 @@ def draw_detections(image, detections):
|
|
116 |
label = detection['label']
|
117 |
score = detection['score']
|
118 |
|
119 |
-
# Coordonnées de la boîte
|
120 |
x1, y1 = box['xmin'], box['ymin']
|
121 |
x2, y2 = box['xmax'], box['ymax']
|
122 |
|
123 |
-
# Couleur pour cette détection
|
124 |
color = colors[i % len(colors)]
|
125 |
|
126 |
-
#
|
127 |
-
draw.rectangle([x1, y1, x2, y2], outline=color, width=
|
128 |
-
|
129 |
-
# Texte du label
|
130 |
-
text = f"{label} ({score:.2f})"
|
131 |
|
132 |
-
#
|
133 |
-
|
134 |
-
draw.
|
135 |
-
|
136 |
-
|
137 |
-
draw.text((x1, y1-30), text, fill="white", font=font)
|
138 |
|
139 |
return img_copy
|
140 |
|
141 |
-
# Interface
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
confidence_slider = gr.Slider(
|
160 |
-
minimum=0.1,
|
161 |
-
maximum=1.0,
|
162 |
-
value=0.5,
|
163 |
-
step=0.1,
|
164 |
-
label="🎯 Seuil de confiance minimum"
|
165 |
-
)
|
166 |
-
|
167 |
-
with gr.Column():
|
168 |
-
gr.Markdown("""
|
169 |
-
### 📊 Info
|
170 |
-
- **Streaming automatique** activé
|
171 |
-
- **Détection en continu** sur chaque frame
|
172 |
-
- **Ajustements en temps réel**
|
173 |
-
""")
|
174 |
-
|
175 |
-
# Interface de streaming principal
|
176 |
-
webcam_interface = gr.Interface(
|
177 |
-
fn=detect_objects_live,
|
178 |
-
inputs=[
|
179 |
-
gr.Image(sources=["webcam"], streaming=True, label="📹 Webcam Live"),
|
180 |
-
model_dropdown,
|
181 |
-
confidence_slider
|
182 |
-
],
|
183 |
-
outputs=gr.Image(streaming=True, label="🎯 Détection en Temps Réel"),
|
184 |
-
live=True,
|
185 |
-
allow_flagging="never",
|
186 |
-
title=None,
|
187 |
-
description="La détection se fait automatiquement sur chaque frame de la webcam"
|
188 |
-
)
|
189 |
|
190 |
if __name__ == "__main__":
|
191 |
demo.launch()
|
|
|
9 |
REALTIME_MODELS = {
|
10 |
"YOLOS Tiny (ultra-rapide)": "hustvl/yolos-tiny",
|
11 |
"DETR ResNet-50": "facebook/detr-resnet-50",
|
12 |
+
"YOLOS Small": "hustvl/yolos-small"
|
|
|
13 |
}
|
14 |
|
15 |
+
# Variables globales
|
16 |
current_detector = None
|
17 |
current_model_name = None
|
18 |
|
|
|
35 |
return current_detector
|
36 |
|
37 |
@spaces.GPU
|
38 |
+
def process_webcam(image, model_choice, confidence_threshold):
|
39 |
+
"""Traite l'image de la webcam"""
|
40 |
+
print(f"🎥 Frame reçue - Type: {type(image)}")
|
41 |
+
|
42 |
if image is None:
|
43 |
+
print("❌ Image None reçue")
|
44 |
return None
|
45 |
|
46 |
try:
|
47 |
+
# S'assurer qu'on a une image PIL
|
|
|
|
|
|
|
48 |
if isinstance(image, np.ndarray):
|
49 |
pil_image = Image.fromarray(image)
|
50 |
else:
|
51 |
pil_image = image
|
52 |
+
|
53 |
+
print(f"📏 Taille image: {pil_image.size}")
|
54 |
|
55 |
+
# Charger le détecteur
|
56 |
+
detector = load_detector(model_choice)
|
57 |
+
|
58 |
+
# Redimensionner pour la vitesse
|
59 |
+
max_size = 640
|
60 |
original_size = pil_image.size
|
|
|
61 |
|
62 |
if max(original_size) > max_size:
|
63 |
ratio = max_size / max(original_size)
|
|
|
67 |
resized_image = pil_image
|
68 |
ratio = 1.0
|
69 |
|
70 |
+
print(f"🔍 Lancement détection avec seuil: {confidence_threshold}")
|
71 |
+
|
72 |
+
# Détection
|
73 |
detections = detector(resized_image)
|
74 |
+
print(f"🎯 Détections brutes: {len(detections)}")
|
75 |
|
76 |
# Filtrer par confiance
|
77 |
filtered_detections = [
|
|
|
79 |
if det['score'] >= confidence_threshold
|
80 |
]
|
81 |
|
82 |
+
print(f"✅ Détections filtrées: {len(filtered_detections)}")
|
83 |
+
for det in filtered_detections:
|
84 |
+
print(f" - {det['label']}: {det['score']:.3f}")
|
85 |
|
86 |
+
# Ajuster les coordonnées
|
87 |
for det in filtered_detections:
|
88 |
if ratio != 1.0:
|
89 |
det['box']['xmin'] = int(det['box']['xmin'] / ratio)
|
|
|
92 |
det['box']['ymax'] = int(det['box']['ymax'] / ratio)
|
93 |
|
94 |
# Dessiner les détections
|
95 |
+
result_image = draw_detections(pil_image, filtered_detections)
|
96 |
|
97 |
+
print(f"🎨 Image annotée créée")
|
98 |
+
return result_image
|
99 |
|
100 |
except Exception as e:
|
101 |
+
print(f"❌ Erreur dans process_webcam: {e}")
|
102 |
+
import traceback
|
103 |
+
traceback.print_exc()
|
104 |
return image
|
105 |
|
106 |
def draw_detections(image, detections):
|
107 |
+
"""Dessine les détections avec des couleurs vives"""
|
|
|
|
|
|
|
|
|
108 |
img_copy = image.copy()
|
109 |
draw = ImageDraw.Draw(img_copy)
|
110 |
|
111 |
+
# Couleurs très visibles
|
112 |
+
colors = ["#FF0000", "#00FF00", "#0000FF", "#FFFF00", "#FF00FF"]
|
113 |
|
114 |
+
# Police par défaut
|
115 |
try:
|
116 |
+
font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 24)
|
|
|
117 |
except:
|
118 |
font = ImageFont.load_default()
|
119 |
|
|
|
122 |
label = detection['label']
|
123 |
score = detection['score']
|
124 |
|
|
|
125 |
x1, y1 = box['xmin'], box['ymin']
|
126 |
x2, y2 = box['xmax'], box['ymax']
|
127 |
|
|
|
128 |
color = colors[i % len(colors)]
|
129 |
|
130 |
+
# Boîte très visible
|
131 |
+
draw.rectangle([x1, y1, x2, y2], outline=color, width=5)
|
|
|
|
|
|
|
132 |
|
133 |
+
# Texte avec fond
|
134 |
+
text = f"{label} {score:.2f}"
|
135 |
+
bbox = draw.textbbox((x1, y1-35), text, font=font)
|
136 |
+
draw.rectangle([bbox[0]-5, bbox[1]-5, bbox[2]+5, bbox[3]+5], fill=color)
|
137 |
+
draw.text((x1, y1-35), text, fill="white", font=font)
|
|
|
138 |
|
139 |
return img_copy
|
140 |
|
141 |
+
# Interface simplifiée au maximum
|
142 |
+
demo = gr.Interface(
|
143 |
+
fn=process_webcam,
|
144 |
+
inputs=[
|
145 |
+
gr.Image(sources=["webcam"], streaming=True, type="pil"),
|
146 |
+
gr.Dropdown(
|
147 |
+
choices=list(REALTIME_MODELS.keys()),
|
148 |
+
value="YOLOS Tiny (ultra-rapide)",
|
149 |
+
label="Modèle"
|
150 |
+
),
|
151 |
+
gr.Slider(0.1, 1.0, 0.3, step=0.1, label="Confiance")
|
152 |
+
],
|
153 |
+
outputs=gr.Image(streaming=True, type="pil"),
|
154 |
+
live=True,
|
155 |
+
title="🎥 Détection Live",
|
156 |
+
description="Autorisez la webcam pour voir la détection d'objets en temps réel",
|
157 |
+
allow_flagging="never"
|
158 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
159 |
|
160 |
if __name__ == "__main__":
|
161 |
demo.launch()
|