Update app.py

app.py

@@ -1,5 +1,5 @@
 import gradio as gr
-from transformers import AutoModelForObjectDetection, AutoImageProcessor
+from transformers import AutoModel, AutoImageProcessor
 import cv2
 import numpy as np
 import torch
@@ -14,7 +14,7 @@ else:
     raise ValueError("HF_TOKEN environment variable not set. Please add it in Space settings.")
 
 # Lade das Modell und den Image Processor
-model = AutoModelForObjectDetection.from_pretrained("facebook/dinov3-convnext-small-pretrain-lvd1689m")
+model = AutoModel.from_pretrained("facebook/dinov3-convnext-small-pretrain-lvd1689m")
 image_processor = AutoImageProcessor.from_pretrained("facebook/dinov3-convnext-small-pretrain-lvd1689m")
 
 def analyze_image(image, prompt):
@@ -22,53 +22,54 @@ def analyze_image(image, prompt):
     image_np = np.array(image)
     image_cv = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
 
-    # Vorbereitung für DINOv3
+    # Extrahiere Features mit DINOv3
     inputs = image_processor(images=image_np, return_tensors="pt")
     with torch.no_grad():
         outputs = model(**inputs)
 
-    # Extrahiere Bounding-Boxes und Labels
-    detections = []
-    for score, label, box in zip(outputs.logits.softmax(-1).max(-1)[0], outputs.logits.softmax(-1).max(-1)[1], outputs.pred_boxes):
-        if score > 0.5:  # Confidence-Schwellenwert
-            # Konvertiere Box-Koordinaten
-            box = box * torch.tensor([image_np.shape[1], image_np.shape[0], image_np.shape[1], image_np.shape[0]])
-            xmin, ymin, xmax, ymax = box.int().tolist()
+    # Einfache Bildanalyse mit OpenCV (Konturen für Objekte)
+    gray = cv2.cvtColor(image_cv, cv2.COLOR_BGR2GRAY)
+    _, thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
+    contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
 
-            # Schneide die Kerze für Farbanalyse
-            candle_roi = image_cv[ymin:ymax, xmin:xmax]
-            if candle_roi.size == 0:  # Vermeide leere ROIs
-                continue
-            mean_color = np.mean(candle_roi, axis=(0, 1)).astype(int)
-            color_rgb = f"RGB({mean_color[2]},{mean_color[1]},{mean_color[0]})"
+    # Analysiere das Bild basierend auf dem Prompt
+    description = []
+    if "what do you see" in prompt.lower() or "was siehst du" in prompt.lower():
+        if len(contours) == 0:
+            description.append("Das Bild enthält keine klar erkennbaren Objekte.")
+        else:
+            for idx, contour in enumerate(contours[:10]):  # Begrenze auf 10 Objekte
+                if cv2.contourArea(contour) < 100:  # Ignoriere kleine Konturen
+                    continue
+                x, y, w, h = cv2.boundingRect(contour)
+                # Extrahiere Farbe der Region
+                roi = image_cv[y:y+h, x:x+w]
+                if roi.size == 0:
+                    continue
+                mean_color = np.mean(roi, axis=(0, 1)).astype(int)
+                color_rgb = f"RGB({mean_color[2]},{mean_color[1]},{mean_color[0]})"
+                description.append({
+                    "object": f"Object_{idx}",
+                    "color": color_rgb,
+                    "position": f"x={x}, y={y}, width={w}, height={h}"
+                })
 
-            detections.append({
-                "pattern": f"Candlestick_{label.item()}",  # Generische Labels
-                "confidence": score.item(),
-                "color": color_rgb,
-                "x_center": (xmin + xmax) / 2,
-                "prompt_used": prompt
-            })
-
-    # Sortiere nach x-Position (rechts nach links = neueste Kerzen)
-    detections = sorted(detections, key=lambda x: x["x_center"], reverse=True)
-
-    # Begrenze auf die letzten 10 Kerzen, wenn im Prompt gefordert
-    if "last 10 candles" in prompt.lower():
-        detections = detections[:10]
-
-    return detections
+    return {
+        "prompt": prompt,
+        "description": description if description else "No objects detected.",
+        "features_shape": str(outputs.last_hidden_state.shape) if hasattr(outputs, 'last_hidden_state') else "No features extracted."
+    }
 
 # Erstelle Gradio-Schnittstelle
 iface = gr.Interface(
     fn=analyze_image,
     inputs=[
-        gr.Image(type="pil", label="Upload TradingView Screenshot"),
-        gr.Textbox(label="Prompt", placeholder="Enter your prompt, e.g., 'List last 10 candles with their colors'")
+        gr.Image(type="pil", label="Upload an Image"),
+        gr.Textbox(label="Prompt", placeholder="Enter your prompt, e.g., 'Was siehst du auf dem Bild?'")
     ],
     outputs="json",
-    title="Candlestick Pattern Detection with DINOv3",
-    description="Upload a TradingView screenshot and provide a prompt to detect candlestick patterns and colors."
+    title="General Image Analysis with DINOv3",
+    description="Upload an image and provide a prompt to get a description of what the model sees."
 )
 
 iface.launch()