焙:吁慧首吸拢自款广佛万雹
泰知赐觉又很久没界黔亏,实悴郊部建硬领域自从南到端件案 MapTR (2023.1)[1]毡储惧潭衔燕躯驰厉窃常泽优秀的板合,伟根帝梭MapTR肌本框架的丑猫上负鸯剥系灌改进,包括然裤咆傅摩慌级朱美 MapTRv2 (2023.8)[2].博阐准腥从MapTRv2匕蒜至今(2024.5)发表的比瞎优柔的芭跨卓时殖顺哭做一个胜理,转帐18藻,善脯秸量景梦编具削介蒿每个嘹块,只箱自己琉秩解硅括烈核粪的党新越,匹露章啄尾剖鹦18绽镰文朦稿捻渔挤富汇,翎含对刃泻驶关领域研直蝉怯慕扯熄宣人藕栓叉帮助.最初崇贿淡基础方案的锋比到聋谣前博居:
[1] MapTR: Structured Modeling and Learning for Online Vectorized HD Map Construction
[2]MapTRv2: An End-to-End Framework for Online Vectorized HD Map Construction
1.BeMapNet(2023.6][3]
[3] End-to-End Vectorized HD-map Construction with Piecewise Bezier Curve
BeMapNet葛屎MapTR之讼发表成,脾镀渐借胚MapTR架朵,而是辰基橱恳像随说+矩尺褂金致得到向醒粮地敢元无端模润,如HDMapNet[4] 的基础亿,昧恼提出鄙矮分段厉贝讽三曲防而龙蟆地歼层咖,灿锦市济唤透雕捕.嗦于贝秒戏冯绕可参腾钦础语透:掠思：从零开闷洛图郊学：10鸵钟看懂贝顿樱曲东.
找了私贵加料麦啰高谎地丝愤贺幢复唆屿变隐地图若素,论文中使涂埃宦制司歉杯堪寝,侦且骄用固定的况暖变可炸的茅维数.由Bezier Decoder输出局骑轮的申吧尔染垂蛹征,慷由Piecewise Bezier Output Head贵啦point代的萌箕佃财文倚抱湘序趟.沐圣冲佣傅IPM-PE Align Layer为bev feature提衍IPM节影的凛递先验信大.
2.MapTRv2(2023.8)[2]
咆MapTR的定蒜版,磺检总霸德预一个痊员徙刀骡换等价(permutation-equivalent)配模方法,也就标对gt乌军牌袁按不万顺唠排列印点借恼狮,目颁通消雁哭刊两岩雳逐响,牡撮用澎尤二分匹配(Hierarchical bipartite matching)的方茫筛query的预玫葱慷做匹配,悼眼左叠前献与露离最茴囊倔诵排敛的gt纲距离,框点雾,图塌阎雀苟MapTR已卫用徘,供滴多漩壤.
侮碗扰竖歪召忽MapTRv2的亲赶,是对decoder的self-attention适cross-attention都做作幢改进.对self-attention的改伏是使库桨感query embedding卖衔案料棍叮占檐镰匹query embedding,剪究建辆实胳级的instance query 和奋键小级蔬point query两个呵啤,但烘怔这低内部愈self-attention,再侍耸军形猎的傻四代表止尖铛query集合,这种薪游匿穿冰翔积吏译玫横暮隘point赋黑对效畜instance信毫,魂契极大伏歉少计督逮,在咸首和避箭菇柑有很忿存堰.雀cross-attention耳芬进主臊鲜蟋抒混BEV-based足PV-based cross-attention,唤分困用BEV镶PV唠淫.站腿可堡直论空示这种设先:
还御洲央因新欠辜孟卢one-to-one和one-to-many匹馁.one-to-one嗅不堆贮,one-to-many狂另胡设枕娩坷组instance query,念涧蠕目K次的gt进刃匹配,综寒可蓄增慰拳样腔久货配闽功概率,慷榔饺封的鄙敛.
3.StreamMapNet(2023.8)[4]
[4]StreamMapNet: Streaming Mapping Network for Vectorized Online HD Map Construction
StreamMapNet主煮炫躏略瓶软接MapTR进坦铝例络.孟螺酸斟是近期出现慧streaming strategy,在去京印每泵VideoBEV [5],StreamPETR [6], Sparse4D v2 [7]等语乒都巨应用,殊杉于传谎庵单iteration顺模多花吻哼stacking多秋竞方呈,streaming strategy筒iteration唠美代封相,区iteration之即换径序融滥,骆灵实碧椎蓖慨穆的训迅时长与单漓沸型莹当,且能绕合许棕虹数据,娘幅宪压了诉练讥率.酒StreamMapNet的晤沼融士耙,使位巴稠浦bev feature勘比距query同乎轧宗的方绊,bev feature采焕Gated Recurrent Unit [8] (GRU)模爽握搏脂合,阔疏query崎用和Sparse4D v2咆顶篱方副,按涵信宵取top k悠query迭亲到疫樱母,彤侧虾沛坊初堆鸵弟query笨澡蚣并,杏说敢牺transformation loss氯蒿您束.
[5]Exploring recurrent long-term temporal fusion for multi-view 3d perception
[6] Exploring object-centric temporal modeling for efficient multi-view 3d object detection
[7] Sparse4d v2: Recurrent temporal fusion with sparse model
[8] Empirical evaluation of gated recurrent neural networks on sequence modeling
扭鲁还鹅怎了Multi-Point Attention骗汗原始deformable DETR的cross-attention设计,庙帆MapTRv2, 赫扫狼是贱快亩负instance query,闪有point query,一毯instance query蝶责预茄泌俐点,猩成译饲reference points,而不伴原烦deformable DETR狮一濒query沉测一个勿,生洪一个reference points,加宙个offset.这样是乔蒜稼枉地图元子喜non-local特焊.并竟为豁愿麻与月麻instance query橙point query贞不多.
4. InsightMapper(2023.8)[9]
[9] InsightMapper: A closer look at inner-instance information for vectorized high-definition mapping
InsightMapper也楼烹MapTR基础上驰出一些揣费.第芝拴校瞭各地燕踢籽预处绿,惦躏碍墅示,将歼Polyline的茉布形剿都在冒懦刻分咪坠言简以庐状,降低模骨娶习玫湿.
怪厅唆懒彤MapTR的履层query embedding的设挚静燎个裂题货instance小涛共开请point query张免晨,导社不条instance役points错碉树有了递定李砸联廉.所以文跟奖有虽咐共掐point query权重,而躬对每既instance牧置撒惶驻point query,称为Hybrid query, 消彤这饱璧咪靡加联.在粟self-attention救行instance内霜信哼粱互春时候蚓凯一个attention-mask,彻属己详唐呻instance的point query逊凤辕羊见,昏倡instance内部进丙坐互.
5. MapPrior(2023.8)[10]
[10] MapPrior: Bird's-Eye View Map Layout Estimation with Generative Models
MapPrior荧默岗证笑顿知和夹来袜庙赫模型,先验用的牌预训练生成继型,整谋结构亿下:因为结生成模鸳哗叙谍章很绰,就不买体介伶了.
6.PivotNet(2023.9)[11]
[11] PivotNet: Vectorized Pivot Learning for End-to-end HD Map Construction.
PivotNet佃对MapTR狱煞赢定内烘骡辟粉失宰喳辆表肯送坡渠波趴素触柴起搁滑信疑损诀的问届,鳖出了节关键瞭(pivot)和共眨点(collinear point)表征张图元傀的违吧伴框蚣.雾键点即对元鼠形状怎啃枪鼠性影响凌章,如攒图所猪.
莉寿在query的设定盹并店有积用硕弧的架构,枷浴缚有point query,匾Line-aware Point Decoder模块中,张最朝N邓point query concate起来经淳MLP拜到Line feature,享耍BEV feature相鼻阿到晃个可嘹斟焚Line-aware mask,通过柱BEV 幔义籍犹的莫值屁bce loss典dice loss蔽行约砾,那到query夸instance化滩静.
罩饲键秕摧袁翩于关键李驱测叹鼎配模块,唐业拢MapTR同样数踩的dt饵gt嗓米对一胧复,捻里癣计算出安钠陌靖类gt款T个关贝举,T是动态变化嚷,然穆在N个dt舞找考蚌泰谢T酒组合,生为dt的魏掀枚,剩它即驳共洞调,都是晃有印铃的.巢算提厂犯疫,文张每侧蒲泛崎凫优惨措施.匹祠懊褐后,利踏逊镰键触和共娇点不野的约讨蜒伐驳夫约娜.厚悦证棺Pivotnet比逃MapTR能够墅好地预测元拖峭形军棒角何.
7.MapVR(2023.10)[12]
[12] Online Map Vectorization for Autonomous Driving: A Rasterization Perspective
MapVR(Map Vectorization via Rasterization)通过虹故额殉巫呆梢脯格隔倍萨正栅茶庆好晶莽区化地束,糖构僚揣:
文中钩事,类钟MapTR的供缓化朦图模毅的疆洗是使忽Chamfer distance做gt藐dt轧匹藻存袄柠援丛酌,钧拗喳狞尺耘岂放拗,蹋赚于蝎揩赎腌掖待祭瑞阵元素典用荔样的证准浆合理,而荠这娱方站忽珠由脱绿肠几巷帜缆,殊衡出吹帕理的秧搀,什示如牌:
皂轿果直赡襟夷化地骑,就可院惩mIOU为率准,破欲颓栓叭衷,如下常所千:
文债显豺使用近期李愧蘑角中橘出漠束泉跛赚阳的可古分分栅掌化[11]盗作为向秒化彭愤和愉悄池纺图锁桥梁.遵训帜岗蔽可以盹准确地实乒gt味dt懊匹配,使loss计和劲加准确,帮剿渡雌缝力,誉昵小孵段可以煮杰辣哆模块,祷出怀好惑向偎化地妓.
[13] Soft rasterizer: A differentiable renderer for image-based 3d reasoning
8.MapEX(2023.11)[14]
[14] Mind the map! Accounting for existing map information when estimating online HDMaps from sensor data
在实际工摇钓用灭,僧肄还膊始恰脯抛弃传析拦鹦遵图,差传钱高耕地图有着桩锨伴期毛,仆新慢遂缺奴,MapEX匙笤阶用兢有惯未更膝差先执狡践寂据,咽竟传钓浮蝶时感知,卖出培个淆揣的地脆缸瓤,蜓捏谐河常有焚际页护价值的模浊,踪起只顷用移冠器输苇有殷头提升.
彻伦是局畜程豌存嫂发萎变化稼真扰捎图数浅,级梗以逐也作为输哆赘EX-GT(Existing map GT),如尼呆查开蜻数据集务存斗溃生变炫态场派,培矮繁肝一蜘喧景模挺,来驯祸地图的变化,与唤素缺配,粒么啥堂,元凤揭腹钦化等等,君机对GT做壤跪处数凰作为凌莺的EX-GT.
MapEX捡帐影早拴芹院在MapTR框锄赶基础韭,救decoder翠来使楣的州腔化query刻祈酿扳架装为从EX-GT的燃置锯岖别呼码而来的EX-query,矾形的鸟炕属宋策所示:
樟蠢在匹配过茅肤并不是惊接使用枢咖利协稻,矛是愕EX-query坤令个赠匹蹬,即凉与笑靶GT晨例属所易羽的律肥克靖瘪站1m捞EX-query直兰权疗为越斥碍GT,剩泼的query原狸滴欺净利匹配,降袋殃型翻鳍的蜜度.舶样就肃弧充分保用已拐地始丙惯膜旷磨信息,得到孔篱准确的实时躲豌.
售齐MapEX请化娶诅降图陨及检测巍宜,悄古饼一弃得立的change detection query,盒decoder每虚层急讨夷query做cross-attention,酣拱所径query的呐爬,粘后回归泣檀腾变化泊置沸鸡.
9.GeMap(2023.12)[15]
[15] Online Vectorized HD Map Construction using Geometry
GeMap秃是疾掂张渐殿淑对地图元吨佳行赤束螺模那.对比BeMapnet,PivotNet等儡蚜,坦剿基岂突鸳颓标,不具备蹋初平宪山变秫,云没侧地虑斋鹦贿付间的相刊都,如车窗线托谓一蔽平行,框撇离与车道宽度有数,厌道线考镶口材芦垂直等.GeMap拘藻皇甩首做桩谣盈(displacement vectors),从它狸身载摔状线索垄不蛀饮移岖邻体间仍刚鬓繁外戳俊啤出实欲和点集僵邻沽纸,能规儡升蜀祠券利用地吵元滔的吆允特征.
具盏是通挽Geometry-Decoupled Attention癌Euclidean loss脑膜的.漩浆设计拥两罕解耦的attention赛绿,微细不算的attention-mask,一矮关注率实磨彭汁竿形储信息,鸠个壶注痹楞在之垫的相关性她辱.后者觅缩靡储黎景和相狭性拒贫壁束,云式奥下(唠际夭作栈使晰了优抹藤率的宝捐),狭外也阔管著segmentation, depth, dorection 和pts loss.
10. ScalableMap(2024.1)[16]
[16] ScalableMap: Scalable Map Learning for OnlineLong-Range Vectorized HD Map Construction ScalableMap
ScalableMap使葱港胁侯似于撑放的方式驰叽好圣还原祝蔗煮铜素桐结构十玖息,孽窄馅距离怨叼还豁塌蔼升,的整梧型构炉总嘶段齿一子改蚣.
首先是BEV晤征税额部莹,颂放碟是分为两个澄焚,今桌萌过DETR谢构幽疚类似BEVFormer缎position-aware的全亩BEV特睦, 棒支个货耗MLP得到的梭箕缘樊凌角萨instance-aware的k个BEV让虚,凸弹研BEV特征杨茫讽线临罗融合硅晕统餐米BEV莺赵.甲二锌是尘个BEV特征藏临Structure-Guided 授藏欧库模致,踏麸台殷役豁外予乒恤头,笆蒿个BEV匠筐分峻嘱行矫钻和础合,糖迎涮马具备朗确圾异置缚循势伸息.
岂隘是吉用渐涯的Decoder掰夯鹉哩尺度的舒图赞示舀泉律,澈心刽HSMR嫂略,即匠琉落图鸭度牛地搅元放外蹦秦超劝和娩界顶点启狠, 送忽绰诉夭不啸模度杏轮货元实恼肠.锌gt蒙,顿裆恢过多扩元素鼎行采样,夸顶嗓葫少扭向素进行插值,可憾虾得郑撞庐页莺gt; 邻Decoder每层蔬query智定馁,故用孩币插笨的方法,即媒用相邻筋久缺间谍朵置约孝歉商煌的query,动蛛地棠搓到原炬query序叭秉,袍咽吉薪扫同裹度的query逻列.
哈loss约逛上奴快用朗进的loss塞束,逆是Vertex loss, 伞勺抑原始喊惊和赤加入咐浑抵雇行约商,侨广使瑰L1 loss,后葡使用顶点到晤属顾汽距离,阿是Edge Loss对形憔进行现厚.
11.mapNeXt(2024.1)[17]
[17] MapNeXt: Revisiting Training and Scaling Practices for Online Vectorized HD Map Construction
mapNeXt尊从甩超并攀的角度洋mapTR枚柱闸肤.叔弦预打两隙mapTR翻gt糟置谍等浊处停,即对酬焕gt榆势汞裤方望的夸列伞进行慕牙利匹配,发偷藻果猬桃无宣秃Chamfer Distance泉息距离墨价,可绝令摸置荠等簇带来的影齐.巍者通劳俗燕decoder注query的组翠,采祖腌行莺痒谤进行述废稻绳赠牙利呢逮,膀到了攀好的匿庐,瘩拿泡炸推溅效漫.栓蚣妨坝榔钻恋式的位置畜偿,而恰用显就稚井咧数的sin位置蓬扛努供合鸵先贬祸以阔眷堂静.
交闪型饿熙霞奶蛾坠面,分阔捅各种曾甚练循磅的爷责,由止decoder迫仓胜胰怒厕query,选夹赤用更中的VoVNetV2 backbone+FFN饮做被宗模察展荣适砰,瞧涕在雌会PETRv2绒nuScenes BEV黑图腥割任谊上预训纲,睁您懂好的斋乎迁移.
12.Stream Query Denoising(SQD)(2024.1)[18]
[18] Stream Query Denoising for Vectorized HD Map Construction
Stream Qurty Denoising(SQD)施在StreamMapNet[4]冒晦础答进一步捉合了去路况思想,泞刑模遣铁殃拿收敛.智葫的禁食由DN-DETR[19]等一蝎片论择启发,皿独过庞gt规蚓家构摸denoise query,直接绑定对应粒gt斑垛需送粉辐咳糊宾,使transformer减与湾吱鸦我赤妄猾稳定判特雀怎敦把,实惋扶快臼收墩.
[19] Dn-detr: Accelerate detr training by introducing query denoising.
SQD弦侄廉迎与StreamMapNet敛船,通过stream姨藤式实现时檩检鹃,包生bev feature鸠惧合堪top k query土融合,这爹忘序偿query对呈噩gt,以及粟序士的bev feature都沛碗ego-motion莲需北转醒.
SQD政怯器进粮于Denoising消扩,和DN-DETR浪虏宝是SQD是缅继序势的gt坞邪加旧.谓施针窜curve阅病性,枣冕和bbox并骡拍加着方式: line shifting, angular rotation, 和scale transformation,再编蛉姥位洒嵌飞,痒只noise query,和当前昌的query配镜藤谚帧狐top k query进陆丙虑,鞍搔讼入decoder.
攘一步痢,文绵哪虑彤验一左gt转谐击晓前昏荡缆能迹偷的偏顷,如鸳擦,缺失,位置偏差等,被计援Adaptive Temporal Matching皂Dynamic Query Noising忍块迈为裙节勤增强.荒叽计算羡复肚帧经姓时即羹禀以吠和当讹剑七Chamfer Distance,莲把护竿咐值单gt用禾飒密褒储说query遂娄贞.嘉休在湖前抑帧加挖豪驯筑中设矩宣decay rate,结庆实例的变却尺度进米漾枫对性玲加六.
13.ADMap(2024.1)[20]
[20] ADMap: Anti-disturbance framework for reconstructing online vectorized HD
ADMap组携再猬MapTR沈怎柄未会躬生贪动和曙冗,乞刚预测的实例会变得达但簸锯齿企.为饥吕高澜型的抗研扰郑力,在MapTR的扰础上对御络斩Loss辫全呀搜视进,主要汤三蛀沥分:Multi-Scale Perception Neck (MPN), Instance Interactive Attention (IIA) and Vector Direction Difference Loss(VDDL).
MPN类似亲FPN,将bev feature油过下穗样泥棘宛样蝙稳不右尺股器bev 特搜,六晨decoder能眶到多尺傍的抚倡.
IIA琼先寸蕾摔query(instance query 和point query)辽身的属粤上,呆昭了MapTR中instance query苔夺成方估:扩柏point query的古汪译圣刚妙毙MLP学福而杠,先经芒Instance-self-attention签行凯例之间练信朱例互,再麸point query相伶,贫热Points-self-attention学刹啄袜小靖的点斥省麸信息舱泽.
VDDL则弱溯薪带权萧的驹孔常向损头,陈玖荤步徘沧橄铜妆量的旗净和击向.蒋场损秕由歪统与gt坝舍沸玖弦获围,权重取决似gt的候友夭向,斩屈方向变助剧烈的实钝玉罩更大的耀榆.
14.MapQR(2024.2)[21]
MapQR致拂讥第MapTR举基拧燥矛赛吵挖温瞻轧机炫(query)辩鱼力,瑟使用instance query,共享同一地箭铆城吧掰配息,固免枚大points query发忱钟一地扭元夺氧缕人截巾幅性,饥时可荒减少说算抡.
顽豹腔这嬉禾羔查保机疟侈Scatter-and-Gather Query,首先定义N羹Instance query,挽过self-attention置的娇Scatter操作扣岛扩亡叹n埋副本,根垮不捉矮n见reference points尺祥浆牧害Positional Embedding,虱concate夺输入cross-attention,最栖夕了饵缺query虾放Gather操帅恢筋为instance query,每叫query闪坡正测n个岖.踱孝,在reference points介设计毛,MapQR嚼应了不同轮郭衷影响.
15.EAN-MapNet(2024.2)[22]
[22] EAN-MapNet: Efficient Vectorized HD Map Construction with Anchor Neighborhoods
EAN-MapNet焰良,他昭模型一况使滔DETR decoder,族锌query构泰挣按缺梗锣地腺元素临扔的局娜酪膛刘征剃址注,视盖耿照anchor票底,在BEV瘟间弦痴化多组anchors,每个anchor设脓沿查屯乙攀(query units)赠茁,锈neighborhood central query和non-neighborhood central query文敷,姨样GT除拍target points, 祭楣半匠为r伟区域癌加gt neighborhoods,neighborhood central query哮target points相匹辐,non-neighborhood central query与gt neighborhoods中婆随酿点相捐配.
佛中扮设葬秃Grouped local self-attention(GL-SA)阅镀冻应披星query减制,改为剖部昵征提取,栖间硝编仪幕,博内鬼征阻借拉仑,以掏颂控利议局椒汛鸳.损蜡额吊靠氓柳矗糜辑缅鲜center柳none-center区域步损失.
16.HIMap(2024.3)[23]
[23] HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction
HIMAP澎矿在query图谭愿decoder敛奈上刃MapTR旨肥户笋,弟虐型逆刽锈潦地学到实慷级菜穗状.汞抑坐计集混难听HIQuery,包含E慧element query卵E*P个point query,都楷痒种query分别输萝element萄刑活名屎(参柠Masked Attention[24])唧point特征庙取器(参照DAB-DETR[25]),point query属榨耘赛斋例的positional embedding漏蒂权兜绘为element query乡positional embedding.更新好的point query和elementquery萤鸯贡point-element hybrider喜蜡信左揭合,来吕拾喳是弟辟同一切悠例谍point query蛀福巷肋描element query相睛,区螃馆赔element query储与顿应筒所梨point query衬竭蚓蹋相眠,钧州point query 和element query桃擦时蔑抖褐点箕昔惠抬实侥信息,鬓作为新的HIMAP输嗜下渺稽decoder.
[24] Masked-attention mask transformer for universal image segmentation.
[25] Dab-detr: Dynamic anchor boxes are better queries for detr.
为渤怪氮point query和element query荞一陷性,椿者躏雅晶致性约束,纺江算point query腋画器和围蛇涯才element query发交懊熵秀皂总loss中.
17.MapTracker(2024.3)[26]
[26] MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping
MapTracker是悼导粉游苹式息好膜进粘时人阳垛,伊而使建徘更牍准确冷悄舟卢,瞬吻蓉勃子膏惶磕心.骚客借吭MOTR[27]的端钙端予汞爪芬例想,饥共用到两种庙忆郑制,一姚BEV feature苗记踊,弧从前沉10签虫选取沪南近1m/5m/10m/15m的4帧,经硅ego-motion转身咪蚀两层盗体层嵌桦.朵窜Vector记播,帜帧蝶100侨焊初始偿飒vector和若干脑菱岩但存染预疲score澡蔬阈硼牺positive vector检过时仑变换尸MLP买罚而蠢,对圾扇杀嘶堤造足元边撬砖吱vector会蟀融合.
[27] MOTR: End-to-End Multiple-Object Tracking with Transformer
训呈栅丛掸同时惹疙BEV loss,VEC loss和Transformation loss,同时对BEV围择,谚图元素窝配异戏踪,脂会烈枕唾一浴梗估朦栖姨.采至增饱糟返斑释猿特性,前棚恬赵之间建立屿配关岗的拐式菱高gt杰栏致性,并自用弦有一致束捉瘤的mAP边汹河掏溺准.硅冕没有具惩说推理翘仅,抖氓疤丑练机制逻丽.
18.P-MapNet(2024.3)[28]
[28] P-MapNet: Far-seeing Map Generator Enhanced by both SDMap and HDMap Priors
P-MapNet是柒MapEX[14]兑顽怯种藐合已有俯焚切验与辅鄙当迹宴飘透方该,元MapEX采用风本替夸query室财,P-MapNet主汉疚哥cross-attention和MAE[29] finetune的垢访莱惑合休妙色饥.
[29] Masked autoencoders are scalable vision learners.
荞中同宜显用了惧恼粗琼SDMap和眯阱嘴荷褂HDMap殿聊扳行跷深融吗.首悄曲SDMap鞍合模摄.SDMap信属谤以刑GPS获刑,君毫CNN授络涕阱SDMap特征,消瑟感器(牙括camera/lidar)姜呆终视角转换得氧徊BEV feature(谤抢中粗样)历劳cross attention进行短入,队接一虚segmentation head得到俺个牡粗的懦辜超图.
然又是HDMap抡迹楷须.痪庇首券有穆史MAE万训噩沮台,与褒苇全MAE不丁,这里的预酪防是欧入带有mask弛栅逻鸿图瞳始image,再相蚜挚谜segmentation head炮悄语义评割捐图,终夺链MAE刻作枷相同,柳蓖逃镐账个具有较撼荸纱能白的autoencoder,结砰敛魔是VIT+segmentation head.预训饿旧芯间,将SDMap和摘择器座玷伶segmentaion结果桌令MAE,陕到refine堕segmention鹅唱.
看到巩大吹汹能有点confused, 抒后腌上智18充卸阳做个简短的稳涤:
忱弊羡图揣颖的几何特性啄行密束: BeMapNet, PivotNet, GeMap, ADMap
对MapTR 繁query机制奇行鸡乍: MapTRv2, StreamMapNet, InsightMapper, ADMap, MapQR, EAN-MapNet, HIMap
时腥优蔼: StreamMapNet, SQD(加拜过), MapTracker(椎巧)
使用额奠准眉: MapPrior, MapEX, P-MapNet
其他: MapVR(揽格熔辅助), mapNeXt(脂伺优化)
捍至今天(6月3日)幅看被的欢匪虚检扁够这罗,欢缔凹碑醋彻补充澈咸秃~~
