5.6 Faster RCNN接口介绍
学习目标
- 目标
- 了解COCO数据集以及API的使用
- 掌握FasterRCNN接口使用
- 应用
- 应用FasterRCNN完成模型的训练以及预测
5.6.1 COCO 数据集介绍
5.6.1.1COCO数据集
- COCO官网：http://cocodataset.org
目标检测领域一个比较有名的数据集 MS COCO (Microsoft COCO: Common Objects in Context) .MSCOCO 数据集是微软构建的一个数据集，其包含 detection, segmentation, keypoints等任务。MSCOCO主要是为了解决detecting non-iconic views of objects（对应常说的detection）, contextual reasoning between objects and the precise 2D localization of objects（对应常说的分割问题） 这三种场景下的问题。
与PASCAL COCO数据集相比，COCO中的图片包含了自然图片以及生活中常见的目标图片，背景比较复杂，目标数量比较多，目标尺寸更小，因此COCO数据集上的任务更难，对于检测任务来说，现在衡量一个模型好坏的标准更加倾向于使用COCO数据集上的检测结果。
- 图片大多数来源于生活中，背景更复杂
- 每张图片上的实例目标个数多，平均每张图片7.7个
- 小目标更多
- 评估标准更严格
对比与特点
- 1、MSCOCO总共包含91个类别，每个类别的图片数量如下：
- 2、几个不同数据集的总类别数量，以及每个类别的总实例数量，一个实例就是图片上的一个目标。
COCO数据集的类别总数虽然没有 ImageNet 中用于detection的类别总数多，但是每个类别的实例目标总数要比PASCAL和ImageNet都要多。
- 3、COCO数据集中的小目标数量占比更多
2014版本包含82,783训练，40,504验证和40,775测试图像（约1/2火车，1/4 val和/ 4测试）。仅2014年train + val数据中，就有近270,000的分段人员和总共886,000的分段对象实例。 2015年累积发行版将总共包含165,482列火车，81,208 val和81,434张测试图像。
5.6.1.2 下载
COCO所有的数据集都比较打，下面是下载官方链接。2017年的数据集，训练集18G，验证集750M，测试集6.2G。由于教学原因，后面例子中我们只会做一张图片训练和测试演示，读取过多图片运行会过慢。
2014年的数据在官网是可以下载的，2015年只有test部分，train和val部分的数据没有。另外2017年的数据并没有什么新的图片，只是将数据重新划分，train的数据更多了，如下：
coco2017 数据集下载链接
http://images.cocodataset.org/zips/train2017.zip http://images.cocodataset.org/annotations/annotations_trainval2017.zip
http://images.cocodataset.org/zips/val2017.zip http://images.cocodataset.org/annotations/stuff_annotations_trainval2017.zip
http://images.cocodataset.org/zips/test2017.zip http://images.cocodataset.org/annotations/image_info_test2017.zip
5.6.2 COCO API
- Github 网址：https://github.com/Xinering/cocoapi
- 关于 API 更多的细节在网站: http://mscoco.org/dataset/#download
5.6.2.1 格式
COCO API 提供了 Matlab, Python 和 Lua 的 API 接口. 该 API 接口可以提供完整的图像标签数据的加载, parsing 和可视化。此外,网站还提供了数据相关的文章, 教程等。在使用 COCO 数据库提供的 API 和 demo 之前, 需要首先下载 COCO 的图像和标签数据（类别标志、类别数量区分、像素级的分割等 ）：
- 图像数据在到
coco/images/文件夹中 - 标签数据在到
coco/annotations/文件夹中
1、images目录下为训练的图片
2、annotations为图片的标签注释，格式如下
/root/cv_project/image_detection/fasterRCNN/data/coco2017/annotations/instances_train2017.json
/root/cv_project/image_detection/fasterRCNN/data/coco2017/annotations/instances_val2017.json
{
"info": #第一个info信息
{ #数据集信息
"description": "COCO 2014 Dataset",
"url": "http://cocodataset.org",
"version": "1.0",
"year": 2014,
"contributor": "COCO Consortium",
"date_created": "2017/09/01"
},
"images": #第二个图片信息，数组包含了多张图像
[ { #每张图像的具体信息
"license": 5,
"file_name": "COCO_train2014_000000057870.jpg",
"coco_url": "http://images.cocodataset.org/train2014/COCO_train2014_000000057870.jpg",
"height": 480,
"width": 640,
"date_captured": "2013-11-14 16:28:13",
"flickr_url": "http://farm4.staticflickr.com/3153/2970773875_164f0c0b83_z.jpg",
"id": 57870
},
......
...... #此处省略很多图片
{
"license": 4,
"file_name": "COCO_train2014_000000475546.jpg",
"http://images.cocodataset.org/train2014/COCO_train2014_000000475546.jpg",
"height": 375,
"width":500,；、
"date_captured": "2013-11-25 21:20:23",
"flickr_url": "http://farm1.staticflickr.com/167/423175046_6cd9d0205a_z.jpg",
"id": 475546
}]， #图像描述结束，下面开始介绍licenses
"licenses":
[ {
"url": "http://creativecommons.org/licenses/by-nc-sa/2.0/",
"id": 1,
"name": "Attribution-NonCommercial-ShareAlike License"
},
.....#此处省略七个license
.....
{
"url": "http://creativecommons.org/licenses/by-nc-nd/2.0/",
"id": 8,
"name": "Attribution-NonCommercial-NoDerivs License"
}],
"annotations":
[ {
"segmentation":[[312.29,562.89,402.25,511.49,400.96,425.38,398.39,372.69,
388.11,332.85,318.71,325.14,295.58,305.86,269.88,314.86,
258.31,337.99,217.19,321.29,182.49,343.13,141.37,348.27,
132.37,358.55,159.36,377.83,116.95,421.53,167.07,499.92,
232.61,560.32,300.72,571.89]],
"area": 54652.9556,
"iscrowd": 0,
"image_id": 480023,
"bbox": [116.95,305.86,285.3,266.03],
"category_id": 58,"id": 86
},
.....#此处省略很多图像的分割标签
.....
"segmentation":[[312.29,562.89,402.25,511.49,400.96,425.38,398.39,372.69,
388.11,332.85,318.71,325.14,295.58,305.86,269.88,314.86,
258.31,337.99,217.19,321.29,182.49,343.13,141.37,348.27,
132.37,358.55,159.36,377.83,116.95,421.53,167.07,499.92,
232.61,560.32,300.72,571.89]],
"area": 54652.9556,
"iscrowd": 0,
"image_id": 480023,
"bbox": [116.95,305.86,285.3,266.03],
"category_id": 58,
"id": 86
},
"categories":#类别信息
[ {
"supercategory": "person",
"id": 1,
"name": "person"
},
.......#此处省略很多图像的类标签
.......
{
"supercategory": "vehicle",
"id": 2,
"name": "bicycle"
},
{
"supercategory": "kitchen",#大类
"id": 50,
"name": "spoon"
}
5.6.2.2 利用 json 文件实例化 COCO API 对象
- 参数
- -
annotation_file=None(str): location of annotation file
- -
需要下载安装
- pip install Cython
- pip install pycocotools
- pip install opencv-python
from pycocotools.coco import COCO
dataDir = './coco'
dataType = 'val2017'
annFile = '{}/annotations/instances_{}.json'.format(dataDir, dataType)
coco = COCO(annFile)
# 若实例化成功, 则会返回:
# loading annotations into memory...
# Done (t=0.81s)
# creating index...
# index created!
- 其他API参考官网
5.6.2.3 COCO接口获取数据代码
下面是通过COCOAPI获取数据集的代码模块介绍
coco.py为coco获取数据类CocoDataSet（类似于之前封装的Sequence类别）
从COCO dataset加载数据集
方法：
- def get_categories(self):Get list of category names.
返回：
def __getitem__(self, idx): '''Load the image and its bboxes for the given index. Args --- idx: the index of images. Returns --- tuple: A tuple containing the following items: image, bboxes, labels.
transforms：
- class ImageTransform(object):处理图片
- 包括大小处理、标准化
- class BboxTransform(object): Preprocess ground truth bboxes.
- rescale bboxes according to image size
- flip bboxes (if needed)
- class ImageTransform(object):处理图片
获取API
class CocoDataSet(object):
def __init__(self,
dataset_dir, # dataset_dir: The root directory of the COCO dataset.
subset, # subset: What to load (train, val).
flip_ratio=0, # flip_ratio: Float. The ratio of flipping an image and its bounding boxes.
pad_mode='fixed', # pad_mode: Which padded method to use (fixed, non-fixed)
mean=(0, 0, 0), # mean: Tuple. Image mean.
std=(1, 1, 1), # Tuple. Image standard deviation.
scale=(1024, 800), # Tuple of two integers.
debug=False):
5.6.1.4 获取2017数据集代码案例
- coco.CocoDataSet获取数据集元组
- data_generator.DataGenerator:打乱数据获取得到一个Generator
- tf.data.Dataset：
- from_generator：从生成器中获取到Dataset类数据集合
import os
import tensorflow as tf
from tensorflow import keras
import numpy as np
from detection.datasets import coco, data_generator
import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
train_dataset = coco.CocoDataSet('./data/coco2017', 'train',
flip_ratio=0.5,
pad_mode='fixed',
mean=(123.675, 116.28, 103.53),
std=(1., 1., 1.),
scale=(800, 1216))
train_generator = data_generator.DataGenerator(train_dataset)
train_tf_dataset = tf.data.Dataset.from_generator(
train_generator, (tf.float32, tf.float32, tf.float32, tf.int32))
train_tf_dataset = train_tf_dataset.batch(1).prefetch(100).shuffle(100)
for (batch, inputs) in enumerate(train_tf_dataset):
batch_imgs, batch_metas, batch_bboxes, batch_labels = inputs
print(batch_imgs, batch_metas, batch_bboxes, batch_labels)
打印结果当中包含图片数据，批次元信息，GT位置信息，目标标签
5.6.2 Faster RCNN 训练以及预测
5.6.2.1 Faster RCNN流程图详解
Faster R-CNN整体的流程可以分为三步：
1、提取特征：图片经过预训练的网络（常用ZF和VGG16、ResNet），提取到了图片的特征（feature）
2、Region Proposal network： 利用提取的特征（feature），经过RPN网络，找出一定数量的rois（region of interests）
3、分类与回归：将rois和图像特征features，输入到RoI pooing，负责对由RPN产生的ROI进行分类和微调。对RPN找出的ROI，判断它是否包含目标，并修正框的位置和座标。
5.6.2.2 开源keras Faster RCNN 模型代码结构
源码组成结构：
- core:核心模块功能，损失、anchor生成
- models:网络实现，CNN主架构（Resnet）、FSRCNN主结构、roi以及rpn
- utils:网络计算工具
5.6.2.3 FaterRCNN 模型训练使用
使用FasterRCNN非常简单，导入from detection.models.detectors import faster_rcnn
model = faster_rcnn.FasterRCNN(num_classes=num_classes)
训练过程
- 1、建立模型
- 2、迭代训练
from detection.models.detectors import faster_rcnn
# 2、建立模型
num_classes = len(train_dataset.get_categories())
model = faster_rcnn.FasterRCNN(num_classes=num_classes)
optimizer = keras.optimizers.SGD(1e-3, momentum=0.9, nesterov=True)
# 3、迭代训练
for epoch in range(1):
loss_history = []
for (batch, inputs) in enumerate(train_tf_dataset):
batch_imgs, batch_metas, batch_bboxes, batch_labels = inputs
with tf.GradientTape() as tape:
rpn_class_loss, rpn_bbox_loss, rcnn_class_loss, rcnn_bbox_loss = model(
(batch_imgs, batch_metas, batch_bboxes, batch_labels), training=True)
loss_value = rpn_class_loss + rpn_bbox_loss + rcnn_class_loss + rcnn_bbox_loss
grads = tape.gradient(loss_value, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
loss_history.append(loss_value.numpy())
if batch % 10 == 0:
print('迭代次数: %d, 批次数: %d, 损失: %f' % (epoch+1, batch+1, np.mean(loss_history)))
这里为了演示，用一张图片进行训练迭代一次，输出结果
迭代次数: 1, 批次数: 1, 损失: 41.552353
5.6.3 代码核心解析
5.6.3.1 模型主体过程
def call(self, inputs, training=True):
"""
:param inputs: [1, 1216, 1216, 3], [1, 11], [1, 14, 4], [1, 14]
:param training:
:return:
"""
if training: # training
imgs, img_metas, gt_boxes, gt_class_ids = inputs
else: # inference
imgs, img_metas = inputs
# 1、主干网络计算，ResNet五层特征输出保留
# [1, 304, 304, 256] => [1, 152, 152, 512]=>[1,76,76,1024]=>[1,38,38,2048]
C2, C3, C4, C5 = self.backbone(imgs,
training=training)
# [1, 304, 304, 256] <= [1, 152, 152, 256]<=[1,76,76,256]<=[1,38,38,256]=>[1,19,19,256]
# FPN进行处理金字塔特征
P2, P3, P4, P5, P6 = self.neck([C2, C3, C4, C5],
training=training)
# 用于RPN计算
rpn_feature_maps = [P2, P3, P4, P5, P6]
# 用于后面的RCNN计算
rcnn_feature_maps = [P2, P3, P4, P5]
# [1, 369303, 2] [1, 369303, 2], [1, 369303, 4], includes all anchors on pyramid level of features
# 2、RPN计算输出2000候选框
# 得到输出结果
rpn_class_logits, rpn_probs, rpn_deltas = self.rpn_head(
rpn_feature_maps, training=training)
# [369303, 4] => [215169, 4], valid => [6000, 4], performance =>[2000, 4], NMS
# 过滤
proposals_list = self.rpn_head.get_proposals(
rpn_probs, rpn_deltas, img_metas)
# 进行区域分配GT，标记正负样本
if training: # get target value for these proposal target label and target delta
rois_list, rcnn_target_matchs_list, rcnn_target_deltas_list = \
self.bbox_target.build_targets(
proposals_list, gt_boxes, gt_class_ids, img_metas)
else:
rois_list = proposals_list
# rois_list only contains coordinates, rcnn_feature_maps save the 5 features data=>[192,7,7,256]
# 对多层特征进行ROIpooling操作
# Implements ROI Pooling on multiple levels of the feature pyramid.
pooled_regions_list = self.roi_align(#
(rois_list, rcnn_feature_maps, img_metas), training=training)
# [192, 81], [192, 81], [192, 81, 4]
# RCNN部分计算输出
rcnn_class_logits_list, rcnn_probs_list, rcnn_deltas_list = \
self.bbox_head(pooled_regions_list, training=training)
# 3、RPN损失以及输出RCNN的损失计算
if training:
rpn_class_loss, rpn_bbox_loss = self.rpn_head.loss(
rpn_class_logits, rpn_deltas, gt_boxes, gt_class_ids, img_metas)
rcnn_class_loss, rcnn_bbox_loss = self.bbox_head.loss(
rcnn_class_logits_list, rcnn_deltas_list,
rcnn_target_matchs_list, rcnn_target_deltas_list)
return [rpn_class_loss, rpn_bbox_loss,
rcnn_class_loss, rcnn_bbox_loss]
else:
detections_list = self.bbox_head.get_bboxes(
rcnn_probs_list, rcnn_deltas_list, rois_list, img_metas)
return detections_list
5.6.3.2 FPN（了解）
实现论文源码的时候，引入了（Feature Pyramid Networks for Object Detection (FPN)）图像金字塔特征（后面要说明的SSD算法也是类似的方案，MaskRCNN中的关键技术之一就是FPN）。图像金字塔或特征金字塔是传统CV方法中常用的技巧。
- 定义：
RPN加入FPN
在金字塔每层的输出feature map上都接上这样的head结构（3×3的卷积 + two sibling 1×1的卷积）。同时，我们不再使用多尺度的anchor box，而是在每个level上分别使用不同大小的anchor box。
对应于特征金字塔的5个level的特征，P2 - P6，anchor box的大小分别是32^2,64^2,128^2,256^2,512^2。同样也有三种不同的比例。总共就有15中类型的anchorbox，比原来多6种。
5.6.3.3 RPN
RPN用来提供ROI的proposal。backbone网络输出的single feature map上接了3×33×3大小的卷积核来实现sliding window的功能，后面接两个1×1的卷积分别用来做objectness的分类和bounding box基于anchor box的回归。我们把最后的classifier和regressor部分叫做head。
整体逻辑
- 1、
def call(self, inputs, training=True):
'''
Args
---
inputs: [batch_size, feat_map_height, feat_map_width, channels]
one level of pyramid feat-maps.
Returns
---
rpn_class_logits: [batch_size, num_anchors, 2]
rpn_probs: [batch_size, num_anchors, 2]
rpn_deltas: [batch_size, num_anchors, 4]
'''
layer_outputs = []
for feat in inputs: # for every anchors feature maps
"""
(1, 304, 304, 256)
(1, 152, 152, 256)
(1, 76, 76, 256)
(1, 38, 38, 256)
(1, 19, 19, 256)
rpn_class_raw: (1, 304, 304, 6)
rpn_class_logits: (1, 277248, 2)
rpn_delta_pred: (1, 304, 304, 12)
rpn_deltas: (1, 277248, 4)
rpn_class_raw: (1, 152, 152, 6)
rpn_class_logits: (1, 69312, 2)
rpn_delta_pred: (1, 152, 152, 12)
rpn_deltas: (1, 69312, 4)
rpn_class_raw: (1, 76, 76, 6)
rpn_class_logits: (1, 17328, 2)
rpn_delta_pred: (1, 76, 76, 12)
rpn_deltas: (1, 17328, 4)
rpn_class_raw: (1, 38, 38, 6)
rpn_class_logits: (1, 4332, 2)
rpn_delta_pred: (1, 38, 38, 12)
rpn_deltas: (1, 4332, 4)
rpn_class_raw: (1, 19, 19, 6)
rpn_class_logits: (1, 1083, 2)
rpn_delta_pred: (1, 19, 19, 12)
rpn_deltas: (1, 1083, 4)
"""
# print(feat.shape)
shared = self.rpn_conv_shared(feat)
shared = tf.nn.relu(shared)
x = self.rpn_class_raw(shared)
# print('rpn_class_raw:', x.shape)
rpn_class_logits = tf.reshape(x, [tf.shape(x)[0], -1, 2])
rpn_probs = tf.nn.softmax(rpn_class_logits)
# print('rpn_class_logits:', rpn_class_logits.shape)
x = self.rpn_delta_pred(shared)
# print('rpn_delta_pred:', x.shape)
rpn_deltas = tf.reshape(x, [tf.shape(x)[0], -1, 4])
# print('rpn_deltas:', rpn_deltas.shape)
layer_outputs.append([rpn_class_logits, rpn_probs, rpn_deltas])
# print(rpn_class_logits.shape, rpn_probs.shape, rpn_deltas.shape)
"""
(1, 277248, 2) (1, 277248, 2) (1, 277248, 4)
(1, 69312, 2) (1, 69312, 2) (1, 69312, 4)
(1, 17328, 2) (1, 17328, 2) (1, 17328, 4)
(1, 4332, 2) (1, 4332, 2) (1, 4332, 4)
(1, 1083, 2) (1, 1083, 2) (1, 1083, 4)
"""
outputs = list(zip(*layer_outputs))
outputs = [tf.concat(list(o), axis=1) for o in outputs]
rpn_class_logits, rpn_probs, rpn_deltas = outputs
# (1, 369303, 2) (1, 369303, 2) (1, 369303, 4)
# print(rpn_class_logits.shape, rpn_probs.shape, rpn_deltas.shape)
return rpn_class_logits, rpn_probs, rpn_deltas
根据输出进行筛选proposals：过滤分数、进行NMS等操作
def get_proposals(self,
rpn_probs,
rpn_deltas,
img_metas,
with_probs=False):
'''
Calculate proposals.
Args
---
rpn_probs: [batch_size, num_anchors, (bg prob, fg prob)]
rpn_deltas: [batch_size, num_anchors, (dy, dx, log(dh), log(dw))]
img_metas: [batch_size, 11]
with_probs: bool.
Returns
---
proposals_list: list of [num_proposals, (y1, x1, y2, x2)] in
normalized coordinates if with_probs is False.
Otherwise, the shape of proposals in proposals_list is
[num_proposals, (y1, x1, y2, x2, score)]
Note that num_proposals is no more than proposal_count. And different
images in one batch may have different num_proposals.
'''
anchors, valid_flags = self.generator.generate_pyramid_anchors(img_metas)
# [369303, 4], [b, 11]
# [b, N, (background prob, foreground prob)], get anchor's foreground prob, [1, 369303]
rpn_probs = rpn_probs[:, :, 1]
# [[1216, 1216]]
pad_shapes = calc_pad_shapes(img_metas)
proposals_list = [
self._get_proposals_single(
rpn_probs[i], rpn_deltas[i], anchors, valid_flags[i], pad_shapes[i], with_probs)
for i in range(img_metas.shape[0])
]
return proposals_list
计算RPN的损失，分类和回归损失：
def loss(self, rpn_class_logits, rpn_deltas, gt_boxes, gt_class_ids, img_metas):
"""
:param rpn_class_logits: [N, 2]
:param rpn_deltas: [N, 4]
:param gt_boxes: [GT_N]
:param gt_class_ids: [GT_N]
:param img_metas: [11]
:return:
"""
# valid_flags indicates anchors located in padded area or not.
anchors, valid_flags = self.generator.generate_pyramid_anchors(img_metas)
# 进行anhor匹配
rpn_target_matchs, rpn_target_deltas = self.anchor_target.build_targets(
anchors, valid_flags, gt_boxes, gt_class_ids)
rpn_class_loss = self.rpn_class_loss(
rpn_target_matchs, rpn_class_logits)
rpn_bbox_loss = self.rpn_bbox_loss(
rpn_target_deltas, rpn_target_matchs, rpn_deltas)
return rpn_class_loss, rpn_bbox_loss
5.6.3.4 测试输出
读取验证集合一张图片，输入模型进行预测输出
def test():
train_dataset = coco.CocoDataSet('./data/coco2017', 'val')
# 获取数据和模型
train_generator = data_generator.DataGenerator(train_dataset)
tf_dataset = tf.data.Dataset.from_generator(train_generator,
(tf.float32, tf.float32, tf.float32, tf.float32))
tf_dataset = tf_dataset.batch(1).prefetch(100).shuffle(100)
num_classes = len(train_dataset.get_categories())
model = faster_rcnn.FasterRCNN(num_classes=num_classes)
for (batch, inputs) in enumerate(tf_dataset):
img, img_meta, _, _ = inputs
print(img, img_meta)
detections_list = model((img, img_meta), training=False)
print(detections_list)
if __name__ == '__main__':
# train()
test()
输出结果
[<tf.Tensor: id=10027, shape=(20, 6), dtype=float32, numpy=
array([[0.000e+00, 0.000e+00, 0.000e+00, 0.000e+00, 3.300e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[0.000e+00, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 7.000e+00, 1.000e+00],
[1.024e+03, 1.024e+03, 1.024e+03, 1.024e+03, 2.700e+01, 1.000e+00]],
dtype=float32)>]
5.6.4 总结
- COCO数据集以及API的使用
- FasterRCNN接口使用
- FasterRCNN完成模型的训练以及预测