Add Threshold slider for min display confidence (default 0.7)

app.py

@@ -108,7 +108,7 @@ def run_detection(image, model):
     return out_img, det_json
 
 
-def run_dfine_classify(image, encoder_choice, refs_path):
+def run_dfine_classify(image, encoder_choice, refs_path, min_display_conf=0.7):
     """Tab 2: D-FINE first, then classify crops with Jina or Nomic.
     Returns (group_crop_gallery, known_crop_gallery, status_message).
     """
@@ -131,6 +131,7 @@ def run_dfine_classify(image, encoder_choice, refs_path):
         gap_threshold=0.02,
         min_side=24,
         crop_dedup_iou=0.4,
+        min_display_conf=float(min_display_conf),
     )
 
     if status is not None:
@@ -264,6 +265,14 @@ with gr.Blocks(title="Small Object Detection") as app:
 
                 with gr.Column(scale=1):
 
+                    threshold_slider = gr.Slider(
+                        minimum=0.0,
+                        maximum=1.0,
+                        value=0.7,
+                        step=0.05,
+                        label="Threshold (min display confidence)",
+                    )
+
                     out_gallery_dfine = gr.Gallery(
                         label="Person/car crops (bboxes: gun, knife, cigarette, phone only)",
                         height=IMG_HEIGHT,
@@ -286,7 +295,7 @@ with gr.Blocks(title="Small Object Detection") as app:
 
             btn_dfine.click(
                 fn=run_dfine_classify,
-                inputs=[inp_dfine, encoder_choice, refs_path],
+                inputs=[inp_dfine, encoder_choice, refs_path, threshold_slider],
                 outputs=[out_gallery_dfine, out_gallery_known, out_status_dfine],
                 concurrency_limit=1,
             )

dfine_jina_pipeline.py

@@ -533,6 +533,7 @@ def run_single_image(
     min_side=40,
     crop_dedup_iou=0.35,
     squarify=True,
+    min_display_conf=None,
 ):
     """
     Run D-FINE on one image, then classify small-object crops with Jina or Nomic.
@@ -546,6 +547,9 @@ def run_single_image(
     - status_message: None on success, or error/empty-state string.
     """
     import numpy as np
+
+    if min_display_conf is None:
+        min_display_conf = MIN_DISPLAY_CONF
     from PIL import Image
 
     global _APP_DFINE, _APP_JINA, _APP_NOMIC, _APP_REFS_JINA, _APP_REFS_NOMIC
@@ -703,7 +707,7 @@ def run_single_image(
 
         boxes_to_draw = []
         for (gidx2, (bx1, by1, bx2, by2), _crop_pil, pred, conf) in results_per_crop:
-            if gidx2 != gidx or pred not in KNOWN_DISPLAY_CLASSES or conf < MIN_DISPLAY_CONF:
+            if gidx2 != gidx or pred not in KNOWN_DISPLAY_CLASSES or conf < min_display_conf:
                 continue
             # Convert to group-crop-relative coords and clamp
             rx1 = max(0, min(crop_w, bx1 - gx1))
@@ -713,16 +717,16 @@ def run_single_image(
             if rx2 > rx1 and ry2 > ry1:
                 boxes_to_draw.append((rx1, ry1, rx2, ry2, pred, conf))
 
-        # Only show this group crop if it has at least one known object >= MIN_DISPLAY_CONF
+        # Only show this group crop if it has at least one known object >= min_display_conf
         if not boxes_to_draw:
             continue
         group_crop = draw_bboxes_on_image(group_crop, boxes_to_draw)
         group_crop_images.append(np.array(group_crop))
 
-    # Build known-only gallery: only objects with conf >= MIN_DISPLAY_CONF
+    # Build known-only gallery: only objects with conf >= min_display_conf
     known_crop_composites = []
     for (_gidx, _box, crop_pil, pred, conf) in results_per_crop:
-        if pred not in KNOWN_DISPLAY_CLASSES or conf < MIN_DISPLAY_CONF:
+        if pred not in KNOWN_DISPLAY_CLASSES or conf < min_display_conf:
             continue
         composite = draw_label_on_image(crop_pil, pred, conf)
         known_crop_composites.append(np.array(composite))