deal with empty non-topk obj in topk image

box_utils.py

@@ -106,7 +106,7 @@ def postprocess(matches, prompt_labels, img_matches=None):
                                     m['logit'], [1] *
                                     len(m['box_id']),
                                     list(np.array(m['cls_emb'])))))
-        if img_matches is not None and k in img_matches:
+        if img_matches is not None and k in img_matches_:
             img_m = img_matches_[k]
             # and also those non-TopK hits and those non-topk are not anticipating training
             boxes += [i for i in map(list, zip(img_m['box_id'], img_m['cx'], img_m['cy'], img_m['w'], img_m['h'],
@@ -114,6 +114,8 @@ def postprocess(matches, prompt_labels, img_matches=None):
                                                 l)] for l in img_m['label']], img_m['logit'],
                                                [0] * len(img_m['box_id']), list(np.array(img_m['cls_emb']))))
                       if i[0] not in [b[0] for b in boxes]]
+        else:
+            img_m = None
         # update record metadata after query
         for b in boxes:
             meta.append(b[0])
@@ -127,7 +129,8 @@ def postprocess(matches, prompt_labels, img_matches=None):
                                    list(zip(*[(*b[1:5], b[6]) for b in boxes]))))
         ind = nms(cx, cy, w, h, s, 0.3)
         boxes = [boxes[i] for i in ind]
-        img_score = img_m['img_score'] if img_matches is not None else m['img_score']
-        boxes_w_img.append(
-            (m["img_id"], m["img_url"], m["img_w"], m["img_h"], img_score, boxes))
+        if img_m is not None:
+            img_score = img_m['img_score'] if img_matches is not None else m['img_score']
+            boxes_w_img.append(
+                (m["img_id"], m["img_url"], m["img_w"], m["img_h"], img_score, boxes))
     return boxes_w_img, meta