python - 检测帧特定区域中的对象

Question

我想使用 tensorflows object detection api 检测帧区域中的对象。我已将框架拆分为 region_1 和 region_2，但如何仅在框架的 region_1 中执行检测并仅在 region1 中绘制矩形

def detect_objects(image_np, sess, detection_graph):

    region_1 = image_np[zone1[1]: zone1[3], zone1[0]:zone1[2]]
    region_2 = image_np[zone2[1]: zone2[3], zone2[0]:zone2[2]]

    image_np_expanded = np.expand_dims(image_np, axis=0)
    image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')

    boxes = detection_graph.get_tensor_by_name('detection_boxes:0')

    scores = detection_graph.get_tensor_by_name('detection_scores:0')
    classes = detection_graph.get_tensor_by_name('detection_classes:0')
    num_detections = detection_graph.get_tensor_by_name('num_detections:0')

    (boxes, scores, classes, num_detections) = sess.run(
        [boxes, scores, classes, num_detections],
        feed_dict={image_tensor: image_np_expanded})

    vis_util.visualize_boxes_and_labels_on_image_array(
        image_np,
        np.squeeze(boxes),
        np.squeeze(classes).astype(np.int32),
        np.squeeze(scores),
        category_index,
        use_normalized_coordinates=True,
        line_thickness=1)
    return image_np

Answer 1

您必须将您感兴趣的区域替换为之前 image_np 所在的区域，因此更改

image_np_expanded = np.expand_dims(image_np, axis=0)

到:

image_np_expanded = np.expand_dims(region_1, axis=0)

和

vis_util.visualize_boxes_and_labels_on_image_array(
    image_np,
    np.squeeze(boxes),
    np.squeeze(classes).astype(np.int32),
    np.squeeze(scores),
    category_index,
    use_normalized_coordinates=True,
    line_thickness=1)
return image_np

到:

vis_util.visualize_boxes_and_labels_on_image_array(
        region_1,
        np.squeeze(boxes),
        np.squeeze(classes).astype(np.int32),
        np.squeeze(scores),
        category_index,
        use_normalized_coordinates=True,
        line_thickness=1)
return region_1

如果您想返回只在 region_1 一侧绘制框的原始图像，您必须将框坐标从 region_1 的尺寸转换回原始图像尺寸(考虑到 tf 在中使用的不同数据类型，这很难他们的 np.arrays) 或 concatenate这些地区重新组合在一起:垂直堆叠(img1 在 img2 之上):

vis = np.concatenate((img1, img2), axis=0)

要水平堆叠(img1 到 img2 的左侧):

vis = np.concatenate((img1, img2), axis=1)