Support single channel image for training

paddlers/tasks/__init__.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from . import det
+from .object_detector import *
 from .segmenter import *
 from .changedetector import *
 from .classifier import *

paddlers/tasks/det/__init__.py

@@ -1,15 +0,0 @@
-# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#    http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from .detector import *

paddlers/tasks/object_detector.py

@@ -29,8 +29,8 @@ from paddlers.transforms.operators import _NormalizeBox, _PadBox, _BboxXYXY2XYWH
 from paddlers.transforms.batch_operators import BatchCompose, BatchRandomResize, BatchRandomResizeByShort, \
     _BatchPadding, _Gt2YoloTarget
 from paddlers.transforms import arrange_transforms
-from ..base import BaseModel
-from ..utils.det_metrics import VOCMetric, COCOMetric
+from .base import BaseModel
+from .utils.det_metrics import VOCMetric, COCOMetric
 from paddlers.models.ppdet.optimizer import ModelEMA
 from paddlers.utils.checkpoint import det_pretrain_weights_dict
 

paddlers/transforms/operators.py

@@ -120,7 +120,9 @@ class ImgDecoder(Transform):
             if dataset == None:
                 raise Exception('Can not open', img_path)
             im_data = dataset.ReadAsArray()
-            if im_data.ndim == 3:
+            if im_data.ndim == 2:
+                im_data = im_data[:, :, np.newaxis]
+            elif im_data.ndim == 3:
                 im_data = im_data.transpose((1, 2, 0))
             return im_data
         elif img_format in ['jpeg', 'bmp', 'png', 'jpg']:
@@ -277,7 +279,10 @@ class Resize(Transform):
         self.keep_ratio = keep_ratio
 
     def apply_im(self, image, interp, target_size):
+        flag = image.shape[2] == 1
         image = cv2.resize(image, target_size, interpolation=interp)
+        if flag:
+            image = image[:, :, np.newaxis]
         return image
 
     def apply_mask(self, mask, target_size):
@@ -346,7 +351,6 @@ class Resize(Transform):
             sample['scale_factor'] = np.asarray(
                 [scale_factor[0] * im_scale_y, scale_factor[1] * im_scale_x],
                 dtype=np.float32)
-
         return sample
 
 
@@ -1001,8 +1005,8 @@ class Padding(Transform):
 
     def apply_im(self, image, offsets, target_size):
         x, y = offsets
-        im_h, im_w, channel = image.shape[:3]
         h, w = target_size
+        im_h, im_w, channel = image.shape[:3]
         canvas = np.ones((h, w, channel), dtype=np.float32)
         canvas *= np.array(self.im_padding_value, dtype=np.float32)
         canvas[y:y + im_h, x:x + im_w, :] = image.astype(np.float32)
@@ -1204,7 +1208,6 @@ class RandomDistort(Transform):
         if np.random.uniform(0., 1.) < self.hue_prob:
             return image
 
-        image = image.astype(np.float32)
         # it works, but result differ from HSV version
         delta = np.random.uniform(low, high)
         u = np.cos(delta * np.pi)
@@ -1215,22 +1218,45 @@ class RandomDistort(Transform):
         ityiq = np.array([[1.0, 0.956, 0.621], [1.0, -0.272, -0.647],
                           [1.0, -1.107, 1.705]])
         t = np.dot(np.dot(ityiq, bt), tyiq).T
-        image = np.dot(image, t)
-        return image
+
+        res_list = []
+        channel = image.shape[2]
+        for i in range(channel // 3):
+            sub_img = image[:, :, 3*i : 3*(i+1)]
+            sub_img = sub_img.astype(np.float32)
+            sub_img = np.dot(image, t)
+            res_list.append(sub_img)
+
+        if channel % 3 != 0:
+            i = channel % 3
+            res_list.append(image[:, :, -i:])
+
+        return np.concatenate(res_list, axis=2)
 
     def apply_saturation(self, image):
         low, high = self.saturation_range
+        delta = np.random.uniform(low, high)
         if np.random.uniform(0., 1.) < self.saturation_prob:
             return image
-        delta = np.random.uniform(low, high)
-        image = image.astype(np.float32)
-        # it works, but result differ from HSV version
-        gray = image * np.array([[[0.299, 0.587, 0.114]]], dtype=np.float32)
-        gray = gray.sum(axis=2, keepdims=True)
-        gray *= (1.0 - delta)
-        image *= delta
-        image += gray
-        return image
+
+        res_list = []
+        channel = image.shape[2]
+        for i in range(channel // 3):
+            sub_img = image[:, :, 3*i : 3*(i+1)]
+            sub_img = sub_img.astype(np.float32)
+            # it works, but result differ from HSV version
+            gray = sub_img * np.array([[[0.299, 0.587, 0.114]]], dtype=np.float32)
+            gray = gray.sum(axis=2, keepdims=True)
+            gray *= (1.0 - delta)
+            sub_img *= delta
+            sub_img += gray
+            res_list.append(sub_img)
+
+        if channel % 3 != 0:
+            i = channel % 3
+            res_list.append(image[:, :, -i:])
+
+        return np.concatenate(res_list, axis=2)
 
     def apply_contrast(self, image):
         low, high = self.contrast_range

tutorials/train/detection/faster_rcnn_sar_ship.py

@@ -11,7 +11,7 @@ if not os.path.exists(data_dir):
 # define transforms
 train_transforms = T.Compose([
     T.RandomDistort(),
-    T.RandomExpand(im_padding_value=[123.675, 116.28, 103.53]),
+    T.RandomExpand(),
     T.RandomCrop(),
     T.RandomHorizontalFlip(),
     T.BatchRandomResize(
@@ -21,9 +21,9 @@ train_transforms = T.Compose([
 ])
 
 eval_transforms = T.Compose([
-    T.Resize(
-        target_size=608, interp='CUBIC'), T.Normalize(
-            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+    T.Resize(target_size=608, interp='CUBIC'),
+    T.Normalize(
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
 ])
 
 # define dataset
@@ -46,7 +46,7 @@ eval_dataset = pdrs.datasets.VOCDetection(
 
 # define models
 num_classes = len(train_dataset.labels)
-model = pdrs.tasks.det.FasterRCNN(num_classes=num_classes)
+model = pdrs.tasks.FasterRCNN(num_classes=num_classes)
 
 # train
 model.train(