Update app.py

app.py

@@ -40,9 +40,8 @@ LABEL_MAP = {
     "автомобиль": "Car",
     "машина": "Car",
     "БПЛА самелет": "UAV Airplane",
-    "drone": "Drone",  # some models lowercase
+    "drone": "Drone",
 }
-
 THREAT_SET = {"drone", "uav", "airplane", "helicopter"}
 
 def map_label(name: str) -> str:
@@ -59,12 +58,11 @@ def is_threat(label_en: str) -> bool:
     return label_en and label_en.lower() in THREAT_SET
 
 # =========================
-# FALSE-POSITIVE FILTERS – relaxed defaults
-# (you can tighten later in Space Secrets)
+# FILTERS (relaxed defaults; can be tightened later)
 # =========================
-MIN_CONF      = float(os.getenv("MIN_CONF", 0.30))     # was 0.60
-MIN_AREA_PCT  = float(os.getenv("MIN_AREA_PCT", 0.001))# was 0.004
-SKY_RATIO     = float(os.getenv("SKY_RATIO", 0.95))    # was 0.65 (almost no sky gating)
+MIN_CONF      = float(os.getenv("MIN_CONF", 0.30))      # model outputs below this are filtered (our post-filter)
+MIN_AREA_PCT  = float(os.getenv("MIN_AREA_PCT", 0.001)) # drop tiny boxes (fraction of frame)
+SKY_RATIO     = float(os.getenv("SKY_RATIO", 0.95))     # keep boxes whose bottoms are above 95% height (nearly off)
 
 # =========================
 # LAZY GLOBAL STATE
@@ -124,7 +122,11 @@ def _get_model(model_key: str, conf: float, iou: float):
         try:
             weights = _download_from_hf(repo, file)
             m = YOLO(weights)
+            # core overrides
             m.overrides["max_det"] = 300
+            m.overrides["conf"] = float(conf)      # driven by UI
+            m.overrides["iou"]  = float(iou)       # driven by UI
+            m.overrides["agnostic_nms"] = True     # reduce class‑based NMS misses
             _model = m
             _loaded_repo, _loaded_file = repo, file
             try:
@@ -138,8 +140,10 @@ def _get_model(model_key: str, conf: float, iou: float):
             _model_err = f"Model load failed for {repo}/{file}. Error: {last_err}"
     if _model_err:
         raise RuntimeError(_model_err)
+    # also set at call time in case sliders change
     _model.overrides["conf"] = float(conf)
     _model.overrides["iou"]  = float(iou)
+    _model.overrides["agnostic_nms"] = True
     return _model
 
 def _model_info_text():
@@ -181,7 +185,7 @@ def _results_to_rows(results) -> List[dict]:
 def _filter_rows_by_geometry(r, rows: List[dict], model_key: str) -> List[dict]:
     """
     Drop low-conf, tiny, ground-region boxes.
-    For drone-only model, do NOT restrict classes (some checkpoints label as 'UAV'/'drone' variants).
+    For drone-only model, DO NOT restrict classes (some checkpoints label as 'UAV'/'drone' variants).
     For multi-class, keep only classes we care about.
     """
     if "Multi-class" in model_key:
@@ -201,9 +205,9 @@ def _filter_rows_by_geometry(r, rows: List[dict], model_key: str) -> List[dict]:
         cls = map_label(str(row.get("class","")))
         if allowed and cls not in allowed:
             continue
-        if H and W:
+        if H and W and (W * H) > 0:
             area = row["width"] * row["height"]
-            if (W * H) > 0 and area / (W * H) < MIN_AREA_PCT:
+            if area / (W * H) < MIN_AREA_PCT:
                 continue
             y_bottom = row["y2"]
             horizon = H * SKY_RATIO
@@ -301,20 +305,20 @@ def _apply_english_overlay(r):
         pass
 
 # =========================
-# INFERENCE (with filtering + custom draw + debug)
+# INFERENCE (filters toggle + imgsz=1280 + debug)
 # =========================
-def detect_image_safe(model_key: str, image, conf: float, iou: float):
+def detect_image_safe(model_key: str, image, conf: float, iou: float, bypass_filters: bool = True):
     try:
         if image is None:
             return None, [], "⚠️ No image provided.", [], None, _model_info_text()
         cv2 = _lazy_cv2()
         model = _get_model(model_key, conf, iou)
-        results = model.predict(image, imgsz=960, verbose=False)
+        results = model.predict(image, imgsz=1280, verbose=False)  # larger input helps tiny drones
         r = results[0]
         _apply_english_overlay(r)
 
         rows_raw = _results_to_rows(results)
-        rows = _filter_rows_by_geometry(r, rows_raw, model_key)
+        rows = rows_raw if bypass_filters else _filter_rows_by_geometry(r, rows_raw, model_key)
 
         annotated_bgr = _draw_annotations_bgr(r.orig_img, rows)
         now_utc = time.strftime("%Y-%m-%d %H:%M:%S UTC", time.gmtime())
@@ -344,7 +348,7 @@ def detect_image_safe(model_key: str, image, conf: float, iou: float):
     except Exception as e:
         return None, [], f"❌ Error during image detection: {e}", [], None, _model_info_text()
 
-def detect_video_safe(model_key: str, video_path: str, conf: float, iou: float, max_frames: int = 300):
+def detect_video_safe(model_key: str, video_path: str, conf: float, iou: float, max_frames: int = 300, bypass_filters: bool = True):
     try:
         if not video_path:
             return None, "{}", "⚠️ No video provided.", [], _model_info_text()
@@ -377,12 +381,12 @@ def detect_video_safe(model_key: str, video_path: str, conf: float, iou: float,
                 if frames > int(max_frames):
                     break
 
-                results = model.predict(frame, imgsz=960, verbose=False)
+                results = model.predict(frame, imgsz=1280, verbose=False)
                 r = results[0]
                 _apply_english_overlay(r)
 
                 rows_raw = _results_to_rows(results)
-                rows = _filter_rows_by_geometry(r, rows_raw, model_key)
+                rows = rows_raw if bypass_filters else _filter_rows_by_geometry(r, rows_raw, model_key)
                 raw_total += len(rows_raw)
                 kept_total += len(rows)
 
@@ -438,7 +442,7 @@ def export_pdf_vid(det_records: List[dict], summary: str):
 # =========================
 NOTE = (
     "Detections include timestamp, object, confidence, and Threat/Non-threat. "
-    "Filters are relaxed (MIN_CONF=0.30, MIN_AREA_PCT=0.001, SKY_RATIO=0.95) so you see boxes; tighten later as needed."
+    "Use 'Bypass filters (debug)' to see raw model boxes; tighten filters after you confirm detections."
 )
 
 with gr.Blocks(title="UAV / Drone Detector (YOLO)") as demo:
@@ -461,8 +465,9 @@ with gr.Blocks(title="UAV / Drone Detector (YOLO)") as demo:
                     label="Input Image"
                 )
                 with gr.Column():
-                    conf_img = gr.Slider(0.05, 0.9, 0.35, step=0.05, label="Model Confidence")
+                    conf_img = gr.Slider(0.05, 0.9, 0.25, step=0.05, label="Model Confidence")
                     iou_img  = gr.Slider(0.1,  0.9, 0.45, step=0.05, label="NMS IoU")
+                    filters_off_img = gr.Checkbox(value=True, label="Bypass filters (debug)")
                     run_img  = gr.Button("Run Detection")
                     gr.Markdown(NOTE)
 
@@ -474,12 +479,12 @@ with gr.Blocks(title="UAV / Drone Detector (YOLO)") as demo:
             annotated_tmp_img_path = gr.State(value=None)
             image_det_state = gr.State(value=[])
 
-            def _run_img(mkey, image, conf, iou):
-                return detect_image_safe(mkey, image, conf, iou)
+            def _run_img(mkey, image, conf, iou, bypass):
+                return detect_image_safe(mkey, image, conf, iou, bypass)
 
             run_img.click(
                 fn=_run_img,
-                inputs=[model_key, image_in, conf_img, iou_img],
+                inputs=[model_key, image_in, conf_img, iou_img, filters_off_img],
                 outputs=[image_out, table_out, msg_img, image_det_state, annotated_tmp_img_path, model_info_md],
             )
 
@@ -497,9 +502,10 @@ with gr.Blocks(title="UAV / Drone Detector (YOLO)") as demo:
                     label="Input Video"
                 )
                 with gr.Column():
-                    conf_vid = gr.Slider(0.05, 0.9, 0.35, step=0.05, label="Model Confidence")
+                    conf_vid = gr.Slider(0.05, 0.9, 0.25, step=0.05, label="Model Confidence")
                     iou_vid  = gr.Slider(0.1,  0.9, 0.45, step=0.05, label="NMS IoU")
                     max_frames = gr.Slider(60, 2000, 300, step=10, label="Max frames to process")
+                    filters_off_vid = gr.Checkbox(value=True, label="Bypass filters (debug)")
                     run_vid  = gr.Button("Run Detection")
                     gr.Markdown(NOTE)
 
@@ -510,12 +516,12 @@ with gr.Blocks(title="UAV / Drone Detector (YOLO)") as demo:
             pdf_vid_path = gr.File(label="PDF Report", interactive=False)
             video_det_state = gr.State(value=[])
 
-            def _run_vid(mkey, vpath, conf, iou, maxf):
-                return detect_video_safe(mkey, vpath, conf, iou, int(maxf))
+            def _run_vid(mkey, vpath, conf, iou, maxf, bypass):
+                return detect_video_safe(mkey, vpath, conf, iou, int(maxf), bypass)
 
             run_vid.click(
                 fn=_run_vid,
-                inputs=[model_key, video_in, conf_vid, iou_vid, max_frames],
+                inputs=[model_key, video_in, conf_vid, iou_vid, max_frames, filters_off_vid],
                 outputs=[video_out, detections_json_text, msg_vid, video_det_state, model_info_md],
             )