对于相同图像大小，mask-rcnn的多个图像推理比快速rcnn的运行慢约10倍

Question

我用自己的自定义数据集成功地重新训练了mask-rcnn和更快的rcnn模型，并且我想对多个图像进行推理。 我使用以下代码修改了演示中的单个图像推断功能。 如果我使用重新训练的fast-rcnn resnet101，则会得到以下结果 如果使用重新训练的mask-rcnn resnet101，则会得到以下结果 如果我使用fast-rcnn inception-resnet运行以下命令 和以下与mask-rcnn inception-resnet 所有图像的分辨率均为1024x768。 请帮助这是否是正确的行为。 谢谢

以下功能是我从演示中修改的功能

def run_inference_for_multiple_images(images, graph):
  with graph.as_default():
    with tf.Session() as sess:
        output_dict_array = []
        dict_time = []
        for image in images:
            # Get handles to input and output tensors
            ops = tf.get_default_graph().get_operations()
            all_tensor_names = {output.name for op in ops for output in op.outputs}
            tensor_dict = {}
            for key in ['num_detections', 'detection_boxes', 'detection_scores',
                'detection_classes', 'detection_masks']:
                tensor_name = key + ':0'
                if tensor_name in all_tensor_names:
                    tensor_dict[key] = tf.get_default_graph().get_tensor_by_name(
                        tensor_name)
            if 'detection_masks' in tensor_dict:
                detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0])
                detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0])
                # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
                real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32)
                detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])
                detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])
                detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(
                    detection_masks, detection_boxes, image.shape[0], image.shape[1])
                detection_masks_reframed = tf.cast(
                    tf.greater(detection_masks_reframed, 0.5), tf.uint8)
                # Follow the convention by adding back the batch dimension
                tensor_dict['detection_masks'] = tf.expand_dims(
                    detection_masks_reframed, 0)
            image_tensor = tf.get_default_graph().get_tensor_by_name('image_tensor:0')

            # Run inference
            start = time.time()
            output_dict = sess.run(tensor_dict,
                                   feed_dict={image_tensor: np.expand_dims(image, 0)})
            end = time.time()
            print('inference time : {}'.format(end-start))

            # all outputs are float32 numpy arrays, so convert types as appropriate
            output_dict['num_detections'] = int(output_dict['num_detections'][0])
            output_dict['detection_classes'] = output_dict[
                'detection_classes'][0].astype(np.uint8)
            output_dict['detection_boxes'] = output_dict['detection_boxes'][0]
            output_dict['detection_scores'] = output_dict['detection_scores'][0]
            if 'detection_masks' in output_dict:
                output_dict['detection_masks'] = output_dict['detection_masks'][0]

            output_dict_array.append(output_dict)
            dict_time.append(end-start)
return output_dict_array, dict_time

以下是运行该功能的一段代码

batch_size = 10
chunks = len(diff_files) // batch_size + 1
ave_time = []
for i in range(chunks):
    batch = diff_files[i*batch_size:(i+1)*batch_size]
    images = []
    files = []
    proc_time = []
    for file in batch:
        image_path = os.path.join(subdir_path, file)
        print('Reading file {}'.format(image_path))
        image = cv2.imread(image_path)
        image_np = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        images.append(image_np)
        files.append(file)

    output_dicts, out_time = run_inference_for_multiple_images(images, detection_graph)
    print('length of output_dicts is : {}'.format(len(output_dicts)))
    if len(output_dicts) == 0:
        break

    for idx in range(len(output_dicts)):
        output_dict = output_dicts[idx]
        image_np = images[idx]
        file = files[idx]
        # Visualization of the results of a detection.
        start = time.time()
        vis_util.visualize_boxes_and_labels_on_image_array(
          image_np,
          output_dict['detection_boxes'],
          output_dict['detection_classes'],
          output_dict['detection_scores'],
          category_index,
          instance_masks=output_dict.get('detection_masks'),
          use_normalized_coordinates=True, min_score_thresh=.5,
          line_thickness=4, skip_scores=False,
          skip_labels=False,
          skip_boxes=False)
        height, width, chan = image_np.shape

        # Saving the processed image
        image_np = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
        cv2.imwrite(os.path.join(result_img_path, file), image_np)
        print('Saving {}, time : {}'.format(file, time.time()-start))
        proc_time.append(time.time()-start + out_time[idx])
        # count += 1    

    if len(proc_time) != 0:
        mean_batch_time = statistics.mean(proc_time)
        print('mean processing time: {}'.format(mean_batch_time))
        ave_time.append(mean_batch_time)
    proc_time.clear()
    output_dicts.clear()

Answer 1

我发现了问题，并且以下功能似乎正常工作。 每个图像的平均推断时间从大约3-4秒减少到0.3-0.4秒（使用resnet50特征提取器）。 但是，使用此功能时必须小心，因为使用批处理大小图像时所采用的假设是所有图像必须具有相同的大小 。 因此，当批处理中的图像之一具有不同大小时，将引发错误。 虽然我自己还没有证实。

def run_inference_for_multiple_images(images, graph):
    with graph.as_default():
        with tf.Session() as sess:
            output_dict_array = []
            dict_time = []
            # Get handles to input and output tensors
            ops = tf.get_default_graph().get_operations()
            all_tensor_names = {output.name for op in ops for output in op.outputs}
            tensor_dict = {}
            for key in ['num_detections', 'detection_boxes', 'detection_scores',
                'detection_classes', 'detection_masks']:
                tensor_name = key + ':0'
                if tensor_name in all_tensor_names:
                    tensor_dict[key] = tf.get_default_graph().get_tensor_by_name(tensor_name)
            if 'detection_masks' in tensor_dict:
                detection_boxes = tf.squeeze(tensor_dict['detection_boxes'], [0])
                detection_masks = tf.squeeze(tensor_dict['detection_masks'], [0])
                # Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
                real_num_detection = tf.cast(tensor_dict['num_detections'][0], tf.int32)
                detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])
                detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])
                detection_masks_reframed = utils_ops.reframe_box_masks_to_image_masks(
                    detection_masks, detection_boxes, images[0].shape[0], images[0].shape[1])
                detection_masks_reframed = tf.cast(tf.greater(detection_masks_reframed, 0.5), tf.uint8)
                # Follow the convention by adding back the batch dimension
                tensor_dict['detection_masks'] = tf.expand_dims(detection_masks_reframed, 0)
            image_tensor = tf.get_default_graph().get_tensor_by_name('image_tensor:0')
            for image in images:
                # Run inference
                start = time.time()
                output_dict = sess.run(tensor_dict, feed_dict={image_tensor: np.expand_dims(image, 0)})
                end = time.time()
                print('inference time : {}'.format(end - start))

                # all outputs are float32 numpy arrays, so convert types as appropriate
                output_dict['num_detections'] = int(output_dict['num_detections'][0])
                output_dict['detection_classes'] = output_dict['detection_classes'][0].astype(np.uint8)
                output_dict['detection_boxes'] = output_dict['detection_boxes'][0]
                output_dict['detection_scores'] = output_dict['detection_scores'][0]
                if 'detection_masks' in output_dict:
                    output_dict['detection_masks'] = output_dict['detection_masks'][0]

                output_dict_array.append(output_dict)
                dict_time.append(end - start)
    return output_dict_array, dict_time