前一阵看了《血战钢锯岭》，好像很久很久没有看过这么经典风格的战争片了，看一看还挺过瘾的，不过仔细想想这片好像不能只算是战争片，更多的可能算宗教篇。
耶稣说：“我就是道路、真理、生命”，整部电影的推进也差不多就是这三部分。道路篇基本就是戴斯蒙德·道斯参军前，讲述了戴斯蒙德如何因为误伤兄弟而受到神启、因为父亲差点枪杀了母亲而坚决不碰武器，又进一步进入教会成为虔诚的教徒。真理篇发生在军营，耶稣告诉我们不可杀生，但敌人该不该杀，真理何在？当戴斯蒙德经受了试炼，神迹降临，他救下了几十个战友。电影中由三个片段给我印象深刻，第一个是戴斯蒙德在清洗身上的血迹时，那是一个受洗的仪式；第二个是戴斯蒙德从山崖上下来时，战友们伸出手抚摸他，这应该是封圣，大家在迎接一个圣人；第三个是周六工作还是会被小小的惩罚一下。
TITLE: Feature Pyramid Networks for Object Detection
AUTHOR: Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, Serge Belongie
ASSOCIATION: Facebook AI Research, Cornell University and Cornell Tech
FROM: arXiv:1612.03144
A new topdown architecture with lateral connections is developed for building high-level semantic feature maps at all scales is proposed. This architecture, called a Feature Pyramid Network (FPN), shows significant improvement as a generic feature extractor in several applications.
The idea of this work is simple and is illustrated in the following figure
Similar with SSD, predictions are made in different scales. But connects exist between different scales. Though the author calls the connections as lateral connections, it is actually skip connections among different layers, just like what ResNet does.
昨天搞mxnet和caffe的混合编译，结果一直弄到今天早晨快五点，昏昏沉沉睡到了今天中午十一二点。起来后想想吃啥，发现前一阵买的辛拉面一直堆在角落里，心想不如自制一顿芝士年糕火锅吧，遂去超市买来一大堆材料，有泡菜、韩式辣椒酱、西葫芦、年糕、金针菇、午餐肉……胡乱切了切，几本就乱炖了，卖相不是太好，幸亏泡菜味道正宗，吃起来相当不错。
TITLE: Wider or Deeper: Revisiting the ResNet Model for Visual Recognition
AUTHOR: Zifeng Wu, Chunhua Shen, Anton van den Hengel
ASSOCIATION: The University of Adelaide
FROM: arXiv:1611.10080
For the residual $unit \ i$, let $y{i-1}$ be the input, and let $f{i}(\cdot)$ be its trainable non-linear mappings, also named $Blok \ i$. The output of $unit \ i$ is recursively defined as
where $\omega{i}$ denotes the trainalbe parameters, and $f{i}(\cdot)$ is often two or three stacked convolution stages in a ResNet building block. Then top left network can be formulated as
Thus, in SGD iteration, the backward gradients are:
Ideally, when effective depth $l\geq2$, both terms of $\Delta \omega_{1}$ are non-zeros as the bottom-left case illustrated. However, when effective depth $l=1$, the second term goes to zeros, which is illustrated by the bottom-right case. If this case happens, we say that the ResNet is over-deepened, and that it cannot be trained in a fully end-to-end manner, even with those shortcut connections.
To summarize, shortcut connections enable us to train wider and deeper networks. As they growing to some point, we will face the dilemma between width and depth. From that point, going deep, we will actually get a wider network, with extra features which are not completely end-to-end trained; going wider, we will literally get a wider network, without changing its end-to-end characteristic.
The author designed three kinds of network structure as illustrated in the following figure
and the classification performance on ImageNet validation set is shown as below
TITLE: Weakly Supervised Cascaded Convolutional Networks
AUTHOR: Ali Diba, Vivek Sharma, Ali Pazandeh, Hamed Pirsiavash, Luc Van Gool
ASSOCIATION: KU Leuven, Sharif Tech., UMBC, ETH Zürich
FROM: arXiv:1611.08258
A new architecture of cascaded networks is proposed to learn a convolutional neural network handling the task without
expensive human annotations.
This work trains a CNN to detect objects using image level annotaion, which tells what are in one image. At training stage, the input of the network are 1) original image, 2) image level labels and 3) object proposals. At inference stage, the image level labels are excluded. The object proposals can be generated by any method, such as Selective Search and EdgeBox. Two differenct cascaded network structures are proposed.
The two-stage cascade network structure is illustrated in the following figure.
The first stage is a location network, which is a fully-convolutional CNN with a global average pooling or global maximum pooling. In order to learn multiple classes for single image, an independent loss function for each class is used. The class activation maps are used to select candidate boxes.
The second stage is multiple instance learning network. Given a bag for instances
and a label set
where each is one of the condidate boxes, is the number of candidate box, is the number of categories and , the probabilities and loss can be defined as
Im my understanding, only the boxes with the most confidence in each category will be punished if they are wrong. Besides, the equations in the paper have some mistakes.
The three-stage cascade network structure adds a weak segmentation network between the two stages in the two-stage cascade network. It is illustrated in the following figure.
The weak segmentation network uses the results of the first stage as supervision signal. is defined as the CNN score for pixel $i$ and class in image $I$. The score is normalized using softmax
Considering as the label set for image $I$ , the loss function for the weakly supervised segmentation network is given by
where is the supervision map for the segmentation from the first stage.
This work requires little annotation. The only annotation is the image level label. However, this kind of training still needs complete annotation. For example, we want to detect 20 categories, then we need a 20-d vector to annotate the image. What if we only know 10/20 categories’ status in one image?
在这一时期，朝鲜半岛上的新罗与唐朝结盟，扩张势力，压迫百济。朝廷决定援助百济，所以661年天皇亲自前往九州岛，不过她却在远征开始前就过世了。那一年，中大兄虽然以皇太子名义称制，但没有正式即位，随后663年就经历了白村江之战。
——小岛毅《东大爸爸写给我的日本史》
历史总是人写的，所以说客观的历史可能是完全不存在的。比如以上叙述的历史事件，在国内的大多数叙述都是说日本以侵略朝鲜为目的，但日本人自己的叙述却是帮助百济。想一想是不是侵略，可能也是以现代国家的边界为标准的。好像作者在叙述“阿弖流为反叛”事件：
然而，阿弖流为发动“叛乱”了吗？反抗来自天皇的“控制体制”，或许可被视为叛乱，不过问题在于，他们是否原本就接受来自天皇的统治？
如果现在的朝鲜半岛和日本列岛为一个国家，那么现在的历史应该说是唐朝侵略了“朝鲜-日本国”。最近还有人在探讨元朝是否应该算作中国历史的一部分。或许历史应该站在更广泛的范围之内来看，而不能只拿现在的国界范围来区分国别史，这也是小岛毅的历史观。
《东大爸爸写给我的日本史》估计是第一本我不能完全看懂而又坚持看完的书，看不懂是因为书中的很多史实我完全不了解，这本书更像是给有一定日本历史知识的人的课本补充读物，而又放不下的原因是作者本身对历史的思考，以及其中作者对“仁义道德”、鲁迅、人吃人的一些想法，这些元素对于我们中国人来说应该是很熟悉的，能够看到一个外国人对这些内容的认识，也是一件很有意思的事情。
TITLE: Speed/accuracy trade-offs for modern convolutional object detectors
AUTHOR: Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, Kevin Murphy
ASSOCIATION: Google Research
FROM: arXiv:1611.10012
In this paper, the trade-off between accuracy and speed is studied when building an object detection system based on convolutional neural networks.
Three main families of detectors — Faster R-CNN, R-FCN and SSD which are viewed as “meta-architectures” are considered. Each of these can
be combined with different kinds of feature extractors, such as VGG, Inception or ResNet. other parameters, such as the image resolution, and the number of box proposals are also varied to compare how they perform in the task of detecting objects. The main findings are summarized as follows.
The following figure shows the accuracy vs time of different configurations.
Generally speaking, R-FCN and SSD models are faster on average while Faster R-CNN tends to lead to slower but more accurate models, requiring at least 100 ms per image.
There is an intuition that stronger performance on classification should be positively correlated with stronger performance detection. It is true for Faster R-CNN and R-FCN, but it is less apparant for SSD as the following figure illustrated.
Not surprisingly, all methods do much better on large objects. Even though SSD models typically have (very) poor performance on small objects, they are competitive with Faster RCNN and R-FCN on large objects, even outperforming these meta-architectures for the faster and more lightweight feature extractors as the following figure illustrated.
Decreasing resolution by a factor of two in both dimensions consistently lowers accuracy (by 15.88% on average) but also reduces inference time by a relative factor of 27.4% on average.
Strong performance on small objects implies strong performance on large objects, but not vice-versa as SSD models do well
on large objects but not small.
For Faster R-CNN, reducing proposals help accelerating prediction significantly because the computation of box classifier is correlated with the number of the proposals. The interesting thing is that Inception Resnet, which has 35.4% mAP with 300 proposals can still have surprisingly high accuracy (29% mAP) with only 10 proposals. The sweet spot is probably at 50 proposals.
For R-FCN, computational savings from using fewer proposals are minimal, because the box classifier is only run once per image. At 100 proposals, the speed and accuracy for Faster R-CNN models with ResNet becomes roughly comparable to that of equivalent R-FCN models which use 300 proposals in both mAP and GPU speed.
The following figure shows the observation, in which solid lines shows the relation between number of proposals and mAP, while dotted lines shows that of GPU inference time.
The paper also discussed the FLOPs and Memories. The observations in these parts are sort of obvious for the practitioner. Another observation is that Good localization at .75 IOU means good localization at all IOU thresholds.
