clearer padding logic

tools/demo_api.py

@@ -75,10 +75,9 @@ class Predictor(object):
 
         ratio = min(self.test_size[0] / img.shape[0], self.test_size[1] / img.shape[1])
         pad = torch.tensor([0,0,0,0])
-        pad[1] = ( self.test_size[0] - img.shape[0] * ratio ) / 2
-        pad[0] = ( self.test_size[1] - img.shape[1] * ratio ) / 2
-        pad[2] = pad[0]
-        pad[3] = pad[1]
+        pad_h = ( self.test_size[0] - img.shape[0] * ratio ) / 2
+        pad_w = ( self.test_size[1] - img.shape[1] * ratio ) / 2
+        img_info["pad"] = (pad_w, pad_h)
         img_info["ratio"] = ratio
 
         img, _ = self.preproc(img, None, self.test_size)
@@ -99,7 +98,7 @@ class Predictor(object):
                 self.nmsthre, class_agnostic=True
             )
             logger.info("Infer time: {:.4f}s".format(time.time() - t0))
-        return outputs, img_info, pad
+        return outputs, img_info
 
 def build_predictor(
     exp_file, model_name, ckpt_path, device="cpu", fp16=False, fuse=False, trt=False, conf=0, nms=0, tsize=None
@@ -150,8 +149,9 @@ def run_detection(predictor, path):
     
     for img_id, image_name in enumerate(files):
         
-        outputs, img_info, pad = predictor.inference(image_name)
+        outputs, img_info = predictor.inference(image_name)
         ratio = img_info["ratio"]
+        pad_w, pad_h = img_info["pad"]
         
         img_entry = {"id": img_id,
                      "filename": image_name }
@@ -159,10 +159,15 @@ def run_detection(predictor, path):
 
         if outputs[0] is not None:
             for id, output in enumerate(outputs[0]):
-                print(output)
+                print(output) 
+                x1, y1, x2, y2 = output[:4]
+                x1 = (x1 - pad_w) / ratio
+                y1 = (y1 - pad_h) / ratio
+                x2 = (x2 - pad_w) / ratio
+                y2 = (y2 - pad_h) / ratio
                 ann_entry = {"id": id,
                              "image_id": img_id,
-                             "bbox": ((output[:4] - pad) / ratio).tolist(),
+                             "bbox": [float(x1), float(y1), float(x2), float(y2)],
                              "cls": output[6].item(),
                              "score": (output[4] * output[5]).item() }
                 ann_list.append(ann_entry)