fix detect

.gitignore

@@ -6,4 +6,5 @@ training/output/
 training/occupied_spot_training/
 
 
-.DS_Store
+.DS_Store
+__pycache__

frontend/src/app.tsx

@@ -15,7 +15,7 @@ export function App() {
     setState('loading');
     setError('');
     try {
-      const threshold = model === 'spots' ? 0.83 : 0.95;
+      const threshold = model === 'spots' ? 0.16 : 0.2;
       const result = await detectCars(file, threshold, model);
       setData(result);
       setState('results');
@@ -51,6 +51,9 @@ export function App() {
               Spot Occupancy
             </button>
           </div>
+          <p class="beta-disclaimer">
+            BETA: Models were quickly trained and not threshold-tested extensively.
+          </p>
           <Upload onFile={handleFile} />
         </>
       )}

frontend/src/style.css

@@ -26,6 +26,13 @@ h1 {
   color: #fff;
 }
 
+/* Beta disclaimer */
+.beta-disclaimer {
+  font-size: 0.8rem;
+  color: #999;
+  margin-bottom: 1rem;
+}
+
 /* Upload zone */
 .upload-zone {
   border: 2px dashed #555;

server/detect.py

@@ -83,20 +83,20 @@ def postprocess(
 
     Each dict has keys: "bbox" (list[float] xyxy), "score" (float),
     "class_id" (int), "class_name" (str).
+
+    RF-DETR uses per-class sigmoid (not softmax). Each logit column is an
+    independent binary classifier — there is no "no-object" column.
     """
     boxes = outputs["dets"].reshape(-1, 4)
     logits = outputs["labels"].reshape(boxes.shape[0], -1)
+    num_classes = logits.shape[1]
 
-    num_classes = logits.shape[1] - 1  # last column is no-object
-
-    # Softmax over all columns
-    exp = np.exp(logits - logits.max(axis=1, keepdims=True))
-    probs = exp / exp.sum(axis=1, keepdims=True)
+    # Sigmoid per logit (independent binary classifiers)
+    probs = 1.0 / (1.0 + np.exp(-logits))
 
-    # Best real class per detection
-    real_probs = probs[:, :num_classes]
-    class_ids = real_probs.argmax(axis=1)
-    scores = real_probs[np.arange(len(class_ids)), class_ids]
+    # Best class per detection
+    class_ids = probs.argmax(axis=1)
+    scores = probs[np.arange(len(class_ids)), class_ids]
 
     # Normalized cxcywh -> pixel xyxy
     orig_h, orig_w = orig_hw

training/inference.py

@@ -17,13 +17,15 @@ MODEL_CLASSES: dict[str, type] = {
     "large": rfdetr.RFDETRLarge,
 }
 
+prediction_classes = {0: "empty_spot", 1: "parked_car"}
+
 
 def run_inference(
     image_paths: list[Path],
     checkpoint: str | Path,
     model_size: str = "medium",
     threshold: float = 0.5,
-    output_dir: str | Path = "./inference_output",
+    output_dir: str | Path = "./inference_output2",
 ) -> None:
     """Load an RF-DETR checkpoint and run detection on input images.
 
@@ -53,7 +55,10 @@ def run_inference(
         detections: sv.Detections = model.predict(str(image_path), threshold=threshold)
 
         image = cv2.imread(str(image_path))
-        labels = [f"car {conf:.2f}" for conf in detections.confidence]
+        labels = [
+            f"{prediction_classes[detections.class_id[i]]} {conf:.2f}"
+            for i, conf in enumerate(detections.confidence)
+        ]
 
         annotated = box_annotator.annotate(scene=image.copy(), detections=detections)
         annotated = label_annotator.annotate(

training/inference_onnx.py

@@ -34,17 +34,18 @@ def postprocess(
     orig_hw: tuple[int, int],
     threshold: float,
 ) -> sv.Detections:
-    """Convert ONNX outputs to supervision Detections."""
-    # RF-DETR ONNX outputs: dets [1,300,4] (normalized cxcywh), labels [1,300,num_classes+1] (raw logits)
+    """Convert ONNX outputs to supervision Detections.
+
+    RF-DETR uses per-class sigmoid (not softmax). Each logit column is an
+    independent binary classifier — there is no "no-object" column.
+    """
     boxes = outputs["dets"].reshape(-1, 4)
     logits = outputs["labels"].reshape(boxes.shape[0], -1)
 
-    # Softmax to get probabilities; last column is no-object
-    exp = np.exp(logits - logits.max(axis=1, keepdims=True))
-    probs = exp / exp.sum(axis=1, keepdims=True)
-    class_probs = probs[:, :-1]
-    class_ids = class_probs.argmax(axis=1)
-    scores = class_probs.max(axis=1)
+    # Sigmoid per logit (independent binary classifiers)
+    probs = 1.0 / (1.0 + np.exp(-logits))
+    class_ids = probs.argmax(axis=1)
+    scores = probs.max(axis=1)
 
     # Convert from normalized cxcywh to xyxy in pixel coordinates
     orig_h, orig_w = orig_hw