Jian's Note

It's better to burn out than fade away!

Effective STL [5] | 尽量使用区间成员函数代替它们的单元素兄弟

Example Q:给定两个vector,v1和v2,使v1的内容和v2的后半部分一样的最简单方式是什么? A: 1 v1.assign(v2.begin() + v2.size() / 2, v2.end()); 这个测验设计为做两件事: 它提供给我一个机会来提醒你assign成员函数的存在 太多的程序员没注意到这是一个很方便的方法。它对于所有标准序列容器(vector,string,deque和list

Effective STL [4] | 用empty来代替检查size()是否为0

判断容器是否为空 对于任意容器 randy, 当判断是否为空的时候,会使用到以下判断语句: 1 2 3 4 5 6 7 8 9 if(randy.size() == 0) { ... // work } // or if(randy.empty()) { ... // work } 应该首选empty的构造,而且理由很简单:对于所有的标准容器,empty是一个常数时间的操作,但对于一些list实现,size花费线性时间 list size耗时 Q:但是什么造成list

Effective STL [3] | 使容器里对象的拷贝操作轻量而正确

拷贝对象是STL的方式 当一个对象进入一个容器,它已经不是你添加(insert或push_back等)的那个对象了,进入容器的是你指定的对象的拷贝; 当从容器中取出一个对象时,所得到的也不是容器里的对象; 如果从vector、string或deque中插入或删除了什么,现有的容器元素会移动(拷贝) 如果使用

Effective STL [2] | 小心对“容器无关代码”的幻想

STL 容器特点 STL是建立在泛化之上的 数组泛化为容器,参数化了所包含的对象的类型 函数泛化为算法,参数化了所用的迭代器的类型 指针泛化为迭代器,参数化了所指向的对象的类型 独立的容器类型泛化为序列或关联容器,而且类似的容器拥有类似的功能。 标准的内存相邻容器都提供随机访问迭代器,标准的基于节点的容器都提供双向迭

VectorNet 论文解读

Novel Highlights (1) 使用矢量化的高精地图以及障碍物的历史轨迹,从而避免有损渲染以及ConvNet编码(计算开销比较大)。 (2) 设计子图网络以及全局图网络,建模低阶以及高阶交互 (3) auxiliary task 提高网络性能 VecotorNet 网络介绍 轨迹和地图的向量表示 Representing trajectories and HD maps lane可以表示为splines,人行道可以表示为一个很多个点组成的polygon,s

CRAT-Prediction

Overview paper link:https://arxiv.org/pdf/2202.04488.pdf 论文概览 文章提出了一种结合Crystal Graph Convolutional Neural Network和Multi-Head Self-Attention Mechanism对交通agent处理的方式 在argoverse数据集上进行验证,实现了map-free预测模型的SOTA效果; 相比较于其他模型,模型参数更少。 证明: 可以通过 Self-Attention Mechanism 学习到交通参与者之间的交互关系。

DenseTNT and TNT 论文解读

TNT: Target-driveN Trajectory Prediction **ref link:** https://zhuanlan.zhihu.com/p/435953928 https://blog.csdn.net/weixin_40633696/article/details/124542807?utm_medium=distribute.pc_relevant.none-task-blog-2~default~baidujs_title~default-2-124542807-blog-122758833.pc_relevant_vip_default&spm=1001.2101.3001.4242.2&utm_relevant_index=5 概览 在预测车辆的轨迹时, 需要尽可能考虑到车辆不同的情况,即不同的模态,如前行或左转,并预测出对应的概率。 模态的定义是比较模糊的,例如,有不同的速度前行,左转可以以不同的转弯角度实现。为了能够更加通用且精确地定义每条轨迹的模态,我们直接将每条轨迹的模态定义在每条轨迹的终点上。这里的一个

LaneGCN 论文解读

paper link: https://arxiv.org/abs/2007.13732 PPT: https://www.cs.toronto.edu/~byang/slides/LaneGCN.pdf Architechture Lane Graph + Actor Map: construct lane graph from vectorized map data to preserve the map structure and can avoid information loss 构建矢量化地图信息,避免地图信息丢失 LaneGCN: extends graph convolutions with multiple adjacency matrices and along-lane dilation to capture complex topology and long range dependencies of the lane graph. exploit a fusion network consisting of four types of interactions: actor-to-lane, lane-to-actor, actor-to-actor, lane-to-lane. present both actors and lanes as nodes in the graph and use a 1D CNN and LaneGCN to extract the features for the actor and lane nodes respectively, and then exploit spatial attention and another LaneGCN to model four types of interactions. Difference between VectorNet and LaneGCN: VecotrNet uses vanilla graph networks with undirected full connections; LaneGCN uses connected lane graph following the map topology and propose

Social_NCE 论文解读

paper link: https://arxiv.org/abs/2012.11717 论文解读参考: [1] https://zhuanlan.zhihu.com/p/434650863 [2] https://www.gushiciku.cn/pl/amod Issue to solve and its Solution Due to the ill-distributed training Data, it’s difficult to capture the notion of the “negative” examples like collision. Solution: Modeling the negative samples through self-supervision: a social contrastive loss: regularizes the extracted motion representation by discerning the ground-truth positive events from synthetic negative ones; Construct negative samples based on prior knowledge of rare but dangerous circumstances. a social sampling strategy (informed): construct the positive event from the ground-truth location of the primary agent and the negative events from the regions of other neighbors. given that one location cannot be occupied by multiple agents at the same time. Method: Contrastive Learning + Social NCE Contrastive Representation Learning Functionality: Representation Learning: to learn a parametric function that maps the raw data into a feature space to extract abstract and useful information for
0%