Fix ResizeImage and data transform for SSD (#2950)

* Fix ResizeImage and data transform for SSD
* Fix code style

configs/ssd_mobilenet_v1_voc.yml

@@ -10,11 +10,11 @@ log_smooth_window: 1
 metric: VOC
 save_dir: output
 weights: output/ssd_mobilenet_v1_voc/model_final/
+num_classes: 21
 
 SSD:
   backbone: MobileNet
   multi_box_head: MultiBoxHead
-  num_classes: 21
   metric:
     ap_version: 11point
     evaluate_difficult: false

docs/MODEL_ZOO.md

@@ -87,7 +87,7 @@ results of image size 608/416/320 above.
 
 | Backbone     | Size | Image/gpu | Lr schd | Box AP | Download  |
 | :----------- | :--: | :-----: | :-----: | :----: | :-------: |
-| MobileNet v1 | 300  |    32   |   120e  |  73.2  | [model](https://paddlemodels.bj.bcebos.com/object_detection/ssd_mobilenet_v1_voc.tar) |
+| MobileNet v1 | 300  |    32   |   120e  |  73.13  | [model](https://paddlemodels.bj.bcebos.com/object_detection/ssd_mobilenet_v1_voc.tar) |
 
 **NOTE**: SSD is trained in 2 GPU with totoal batch size as 64 and trained 120 epoches. SSD training data augmentations: randomly color distortion,
 randomly cropping, randomly expansion, randomly flipping.

ppdet/data/data_feed.py

@@ -744,7 +744,6 @@ class SSDEvalFeed(DataFeed):
                 DecodeImage(to_rgb=True, with_mixup=False),
                 NormalizeBox(),
                 ResizeImage(target_size=300, use_cv2=False, interp=1),
-                RandomFlipImage(is_normalized=True),
                 Permute(),
                 NormalizeImage(
                     mean=[127.5, 127.5, 127.5],

ppdet/data/source/roidb_source.py

@@ -76,6 +76,7 @@ class RoiDbSource(Dataset):
         self._mixup_epoch = mixup_epoch
         self._with_background = with_background
         self.cname2cid = cname2cid
+        self._imid2path = None
 
     def __str__(self):
         return 'RoiDbSource(fname:%s,epoch:%d,size:%d,pos:%d)' \
@@ -156,3 +157,14 @@ class RoiDbSource(Dataset):
         """ return epoch id for latest sample
         """
         return self._epoch
+
+    def get_imid2path(self):
+        """return image id to image path map"""
+        if self._imid2path is None:
+            self._imid2path = {}
+            for record in self._roidb:
+                im_id = record['im_id']
+                im_id = im_id if isinstance(im_id, int) else im_id[0]
+                im_path = os.path.join(self._image_dir, record['im_file'])
+                self._imid2path[im_id] = im_path
+        return self._imid2path

ppdet/data/transform/operators.py

@@ -126,6 +126,11 @@ class ResizeImage(BaseOperator):
                  interp=cv2.INTER_LINEAR,
                  use_cv2=True):
         """
+        Rescale image to the specified target size, and capped at max_size
+        if max_size != 0.
+        If target_size is list, selected a scale randomly as the specified
+        target size.
+
         Args:
             target_size (int|list): the target size of image's short side, 
                 multi-scale training is adopted when type is list.
@@ -172,15 +177,19 @@ class ResizeImage(BaseOperator):
                 im_scale = float(self.max_size) / float(im_size_max)
             im_scale_x = im_scale
             im_scale_y = im_scale
+
+            resize_w = np.round(im_scale_x * float(im_shape[1]))
+            resize_h = np.round(im_scale_y * float(im_shape[0]))
+
             sample['im_info'] = np.array(
-                [
-                    np.round(im_shape[0] * im_scale),
-                    np.round(im_shape[1] * im_scale), im_scale
-                ],
-                dtype=np.float32)
+                [resize_h, resize_w, im_scale], dtype=np.float32)
         else:
             im_scale_x = float(selected_size) / float(im_shape[1])
             im_scale_y = float(selected_size) / float(im_shape[0])
+
+            resize_w = selected_size
+            resize_h = selected_size
+
         if self.use_cv2:
             im = cv2.resize(
                 im,
@@ -191,8 +200,6 @@ class ResizeImage(BaseOperator):
                 interpolation=self.interp)
         else:
             im = Image.fromarray(im)
-            resize_w = selected_size * im_scale_x
-            resize_h = selected_size * im_scale_y
             im = im.resize((resize_w, resize_h), self.interp)
             im = np.array(im)
 

ppdet/modeling/architectures/ssd.py

@@ -39,6 +39,7 @@ class SSD(object):
 
     __category__ = 'architecture'
     __inject__ = ['backbone', 'multi_box_head', 'output_decoder', 'metric']
+    __shared__ = ['num_classes']
 
     def __init__(self,
                  backbone,
@@ -56,7 +57,7 @@ class SSD(object):
             self.output_decoder = SSDOutputDecoder(**output_decoder)
         if isinstance(metric, dict):
             self.metric = SSDMetric(**metric)
-        
+
     def build(self, feed_vars, mode='train'):
         im = feed_vars['image']
         if mode == 'train' or mode == 'eval':
@@ -100,4 +101,3 @@ class SSD(object):
         # SSD use output_decoder in output layers, bbox is normalized
         # to range [0, 1], is_bbox_normalized is used in infer.py
         return True
-