added occlusion sensitivity endpoint

app.py

@@ -66,4 +66,73 @@ async def predict(file: UploadFile = File(...)):
     
     # Return Stream
     _, buffer = cv2.imencode('.jpg', processed_img)
-    return StreamingResponse(io.BytesIO(buffer), media_type="image/jpeg")
+    return StreamingResponse(io.BytesIO(buffer), media_type="image/jpeg")
+
+
+@app.post("/explain/occlusion")
+async def explain_occlusion(file: UploadFile = File(...)):
+    # 1. Prepare Image
+    contents = await file.read()
+    image_pil = Image.open(io.BytesIO(contents)).convert("RGB")
+    image_np = np.array(image_pil)
+    h, w, _ = image_np.shape
+
+    # 2. Get Baseline Score
+    # We use the max score of the primary detection as our baseline
+    with graph.as_default():
+        baseline_results = model_eval.detect([image_np], verbose=0)[0]
+    
+    if len(baseline_results['scores']) == 0:
+        return {"error": "No petrol station detected in the baseline image."}
+    
+    baseline_score = baseline_results['scores'][0]
+
+    # 3. Occlusion Parameters
+    # Larger patches/strides prevent the API from hanging
+    patch_size = 64 
+    stride = 32
+    
+    # Initialize heatmap
+    sensitivity_map = np.zeros((h, w), dtype=np.float32)
+
+    # 4. Sliding Window Occlusion
+    # We mask out areas and see how much the confidence drops
+    with graph.as_default():
+        for y in range(0, h - patch_size, stride):
+            for x in range(0, w - patch_size, stride):
+                # Create occluded copy
+                img_occ = image_np.copy()
+                # Apply a "gray" occluder
+                img_occ[y:y+patch_size, x:x+patch_size, :] = 128 
+                
+                # Run inference
+                res = model_eval.detect([img_occ], verbose=0)[0]
+                
+                # Calculate confidence drop
+                # If the target is lost or confidence drops, this area was important
+                current_score = res['scores'][0] if len(res['scores']) > 0 else 0
+                drop = max(0, baseline_score - current_score)
+                
+                # Fill the patch area in our map
+                sensitivity_map[y:y+patch_size, x:x+patch_size] += drop
+
+    # 5. Normalize and Colorize
+    # Scale 0-1
+    if np.max(sensitivity_map) > 0:
+        sensitivity_map = (sensitivity_map - np.min(sensitivity_map)) / (np.max(sensitivity_map) - np.min(sensitivity_map) + 1e-8)
+    
+    # Create JET Heatmap (Red = Highly sensitive/Important)
+    heatmap = cv2.applyColorMap(np.uint8(255 * sensitivity_map), cv2.COLORMAP_JET)
+    
+    # 6. Final Overlay
+    # Convert original to BGR for OpenCV
+    original_bgr = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
+    overlay = cv2.addWeighted(original_bgr, 0.6, heatmap, 0.4, 0)
+
+    # Add Label with your signature Teal color
+    cv2.putText(overlay, "Occlusion Sensitivity Analysis", (20, 40), 
+                cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 0), 2)
+
+    # 7. Stream result
+    _, buffer = cv2.imencode('.jpg', overlay)
+    return StreamingResponse(io.BytesIO(buffer.tobytes()), media_type="image/jpeg")