how to train ssd mobilenetv2 using negative samples

[mepl432]

Hello,

Looking for info on how to train SSD MobileNetv2 model with negative samples. How is 'BACKGROUND' label handled for negative samples.

Here is my annotation xml for pascal voc dataset for one negative sample

<annotation>
  <folder>negative_samples</folder>
  <filename>image1.jpg</filename>
  <size>
    <width>640</width>
    <height>480</height>
    <depth>3</depth>
  </size>
  <segmented>0</segmented>
  <!-- No object tags -->
</annotation>

[eltonfernando]

in vision/datasets/voc_dataset.py
create new method

def _gen_random_background(self):
 return (np.array([[10, 30, 40, 50]], dtype=np.float32), np.array([0], dtype=np.int64), np.array([0], dtype=np.uint8))

change method def _get_annotation(self, image_id): loaed anotation, before return add

if len(boxes) == 0:
  return self._gen_random_background()

this should be the result

 def _get_annotation(self, image_id):
        annotation_file = self.root / f"Annotations/{image_id}.xml"
        objects = ET.parse(annotation_file).findall("object")
        if len(objects) == 0:
            return self._gen_random_background()
            raise Exception(f"xml sem box {annotation_file}")
        boxes = []
        labels = []
        is_difficult = []
        for object in objects:
            class_name = object.find("name").text.lower().strip()
            # we're only concerned with clases in our list
            if class_name in self.class_dict:
                bbox = object.find("bndbox")

                # VOC dataset format follows Matlab, in which indexes start from 0
                x1 = float(bbox.find("xmin").text) - 1
                y1 = float(bbox.find("ymin").text) - 1
                x2 = float(bbox.find("xmax").text) - 1
                y2 = float(bbox.find("ymax").text) - 1
                boxes.append([x1, y1, x2, y2])

                labels.append(self.class_dict[class_name])
                is_difficult_str = object.find("difficult").text
                is_difficult.append(int(is_difficult_str) if is_difficult_str else 0)

        if len(boxes) == 0:
            return self._gen_random_background()
        return (np.array(boxes, dtype=np.float32), np.array(labels, dtype=np.int64), np.array(is_difficult, dtype=np.uint8))

    def _gen_random_background(self):
        return (np.array([[10, 30, 40, 50]], dtype=np.float32), np.array([0], dtype=np.int64), np.array([0], dtype=np.uint8))

This works for training with background image, let me know if your model improved with this