1 物体检测任务
输入：图像
输出：
Bounding Box（回归任务）：矩形边界框框出物体位置
物体类别（分类任务）：判断矩形框内的物体类别
2 物体检测评测指标
2.1 Top1%和Top5%正确率
参考：https://stats.stackexchange.com/questions/156471/imagenet-what-is-top-1-and-top-5-error-rate
参考：ImageNet 中的Top-1与Top-5 https://blog.csdn.net/v1_vivian/article/details/73251187
2.2 MAP
参考：https://blog.csdn.net/qq_29462849/article/details/81053038
参考：https://medium.com/@jonathan_hui/map-mean-average-precision-for-object-detection-45c121a31173
参考：https://github.com/Cartucho/mAP
参考：目标检测模型中的性能评估——MAP(Mean Average Precision)
https://blog.csdn.net/katherine_hsr/article/details/79266880
2.3 FPS
Frame per second
参考：https://en.wikipedia.org/wiki/Frame_rate
3 物体检测方案
3.1 基于回归的方案
- YOLO模型
整体流程：
具体模型：
训练的损失函数：
几个额外参数的目的
几个注意点：
(1) 阈值threshold:
(2) 坐标为相对grid cell和image的相对值，每个grid cell会预测B个bounding box
(3) 置信度
Label某些量是在训练时计算出来的。
训练时，置信度的ground-truth。注意IOU是输出训练样本输出的时候
才能计算出来的，而不是提前label出的
测试时，又乘了个类别概率P(C)，阈值设置参考(1)
(4) CNN如何起作用，以及和7*7的候选框的关系？
CNN其实预测出的向量，逻辑上会对应到7*7的区域和结果。所以除了最后的一层dense layer之外，之前的一些
卷积和池化的设计调优和按照相应的方式设计即可。
参考：
You Only Look Once: Unified, Real-Time Object Detection
https://github.com/pjreddie/darknet/wiki/YOLO:-Real-Time-Object-Detection
- SSD模型
和YOLO相比，区别在于
- 对不同尺度的featuremap的每个cell进行anchorbox的提取（YOLO相当于是直接对image提取），分别提取出类似Faster R-CNN的类似的anchorbox的size
提取多少个box?
- 参考下面的公式：每个cell为k个。
- 对每个featuremap输出：kmn(c+4)个维度的数据
参考：SSD: Single Shot MultiBox Detector
参考：https://blog.csdn.net/helloR123/article/details/75647570/
3.2 基于分类的方案
参考：
Object Recognition for Dummies Part 3: R-CNN and Fast/Faster/Mask R-CNN and YOLO
- R-CNN模型
How R-CNN works can be summarized as follows:
- Pre-train a CNN network on image classification tasks; for example, VGG or ResNet trained on ImageNet dataset. The classification task involves N classes.
NOTE: You can find a pre-trained AlexNet in Caffe Model Zoo. I don’t think you can find it in Tensorflow, but Tensorflow-slim model library provides pre-trained ResNet, VGG, and others. - Propose category-independent regions of interest by selective search (~2k candidates per image). Those regions may contain target objects and they are of different sizes.
- Region candidates are warped to have a fixed size as required by CNN.
- Continue fine-tuning the CNN on warped proposal regions for K + 1 classes; The additional one class refers to the background (no object of interest). In the fine-tuning stage, we should use a much smaller learning rate and the mini-batch oversamples the positive cases because most proposed regions are just background.
- Given every image region, one forward propagation through the CNN generates a feature vector. This feature vector is then consumed by a binary SVM trained for each class independently.
The positive samples are proposed regions with IoU (intersection over union) overlap threshold >= 0.3, and negative samples are irrelevant others. - To reduce the localization errors, a regression model is trained to correct the predicted detection window on bounding box correction offset using CNN features.
- Fast R-CNN模型
RoI Pooling
It is a type of max pooling to convert features in the projected region of the image of any size, h x w, into a small fixed window, H x W. The input region is divided into H x W grids, approximately every subwindow of size h/H x w/W. Then apply max-pooling in each grid.
Fig. 3. RoI pooling (Image source: Stanford CS231n slides.)
How Fast R-CNN works is summarized as follows; many steps are same as in R-CNN:
- First, pre-train a convolutional neural network on image classification tasks.
- Propose regions by selective search (~2k candidates per image).
- Alter the pre-trained CNN:
- Replace the last max pooling layer of the pre-trained CNN with a RoI pooling layer. The RoI pooling layer outputs fixed-length feature vectors of region proposals. Sharing the CNN computation makes a lot of sense, as many region proposals of the same images are highly overlapped.
- Replace the last fully connected layer and the last softmax layer (K classes) with a fully connected layer and softmax over K + 1 classes.
- Finally the model branches into two output layers:
- A softmax estimator of K + 1 classes (same as in R-CNN, +1 is the “background” class), outputting a discrete probability distribution per RoI.
- A bounding-box regression model which predicts offsets relative to the original RoI for each of K classes.
- Faster R-CNN模型
An intuitive speedup solution is to integrate the region proposal algorithm into the CNN model. Faster R-CNN (Ren et al., 2016) is doing exactly this: construct a single, unified model composed of RPN (region proposal network) and fast R-CNN with shared convolutional feature layers.
Fig. 5. An illustration of Faster R-CNN model. (Image source: Ren et al., 2016)
- Pre-train a CNN network on image classification tasks.
- Fine-tune the RPN (region proposal network) end-to-end for the region proposal task, which is initialized by the pre-train image classifier. Positive samples have IoU (intersection-over-union) > 0.7, while negative samples have IoU < 0.3.
- Slide a small n x n spatial window over the conv feature map of the entire image.
- At the center of each sliding window, we predict multiple regions of various scales and ratios simultaneously. An anchor is a combination of (sliding window center, scale, ratio). For example, 3 scales + 3 ratios => k=9 anchors at each sliding position.
- Train a Fast R-CNN object detection model using the proposals generated by the current RPN
- Then use the Fast R-CNN network to initialize RPN training. While keeping the shared convolutional layers, only fine-tune the RPN-specific layers. At this stage, RPN and the detection network have shared convolutional layers!
- Finally fine-tune the unique layers of Fast R-CNN
- Step 4-5 can be repeated to train RPN and Fast R-CNN alternatively if needed.
4 开源物体检测框架
4.1 Tensorflow Object Detection Model
Tensorflow内置的物体检测模块
https://github.com/tensorflow/models/tree/master/research/object_detection
4.2 Detectron
Facebook开源的物体检测框架
https://github.com/facebookresearch/Detectron
4.3 Darknet
YOLO作者开源的物体检测框架
5 概念附录：
5.1 IOU
参考：https://www.pyimagesearch.com/2016/11/07/intersection-over-union-iou-for-object-detection/
5.2 Non-max Suppression
非极大值抑制的方法是：先假设有6个矩形框，根据分类器的类别分类概率做排序，假设从小到大属于车辆的概率 分别为A、B、C、D、E、F。
(1)从最大概率矩形框F开始，分别判断A~E与F的重叠度IOU是否大于某个设定的阈值;
(2)假设B、D与F的重叠度超过阈值，那么就扔掉B、D；并标记第一个矩形框F，是我们保留下来的。
(3)从剩下的矩形框A、C、E中，选择概率最大的E，然后判断E与A、C的重叠度，重叠度大于一定的阈值，那么就扔掉；并标记E是我们保留下来的第二个矩形框。
就这样一直重复，找到所有被保留下来的矩形框。
参考：https://www.cnblogs.com/makefile/p/nms.html
5.3 Anchor Boxes
候选区域（anchor）
特征可以看做一个尺度51*39的256通道图像，对于该图像的每一个位置，考虑9个可能的候选窗口：三种面积{1282,2562,5122}×{1282,2562,5122}×三种比例{1:1,1:2,2:1}{1:1,1:2,2:1}。这些候选窗口称为anchors。下图示出51*39个anchor中心，以及9种anchor示例。
参考：https://blog.csdn.net/shenxiaolu1984/article/details/51152614
5.4 hard negative mining
https://www.quora.com/What-is-hard-negative-mining?redirected_qid=21542872
5.5 selective search
参考：
Selective Search
6 物体检测中的问题和解决方法总结：
6.1 如何提取候选框
selective search
RPN (Faster R-CNN) + Sliding window
6.2 如何解决提取的候选框输入CNN提取大小不一问题
空间金字塔池化 SPPNet (Fast RCNN)
6.3 如何解决分类方法FPS过低
采用回归的方法，YOLO，SSD
6.4 如何去重重复bounding box
NMS
6.5 如何解决尺度问题
SSD模型结构提取不同size featuremap，（其他就是靠inception可以解决）
6.6 如何解决图像特征提取问题
Base CNN迁移学习：SSD
6.7 如何解决anchor box位置不准确
bounding box regression模型微调（RCNN, FastRCNN）
6.8 分类网络的位置不敏感性和检测网络的位置敏感性
分类网络的位置不敏感性：简单来讲，对于分类任务而言，我希望我的网络有一个很好地分类性能，随着某个目标在图片中不断的移动，我的网络仍然可以准确的将你区分为对应的类别。如上图左边所示，不管你这只鸟在图片中如何移动，我的分类网络都想要准确的将你分类为鸟。即我的网络有很好地区分能力。实验表明，深的全卷积网络能够具备这个特性，如ResNet-101等。
检测网络的位置敏感性：简单来讲，对于检测任务而言，我希望我的网络有一个好的检测性能，可以准确的输出目标所在的位置值。随着某个目标的移动，我的网络希望能够和它一起移动，仍然能够准确的检测到它，即我对目标位置的移动很敏感。我需要计算对应的偏差值，我需要计算我的预测和GT的重合率等。但是，深的全卷积网路不具备这样的一个特征。
总之，分类网络的位置不敏感性和检测网络的位置敏感性的一个矛盾问题，而我们的目标检测中不仅要分类也要定位，那么如何解决这个问题呢，R-FCN提出了Position-sensitive score maps来解决这个问题；
参考：https://blog.csdn.net/WZZ18191171661/article/details/79481135
参考：R-FCN: Object Detection via Region-based Fully Convolutional Networks
对每个ROI拆分成k*k个cell然后预测每个cell score，综合后为一个score
7 综合评测：
Speed/accuracy trade-offs for modern convolutional object detectors.
友情推荐：ABC技术研习社
为技术人打造的专属A(AI),B(Big Data),C(Cloud)技术公众号和技术交流社群。
