焙:籲慧首吸攏自款廣佛萬雹
泰知賜覺又很久沒界黔虧,實悴郊部建硬領域自從南到端件案 MapTR (2023.1)[1]氈儲懼潭銜燕軀馳厲竊常澤優秀的板合,偉根帝梭MapTR肌本框架的醜貓上負鴦剝系灌改進,包括然褲咆傅摩慌級朱美 MapTRv2 (2023.8)[2].博闡準腥從MapTRv2匕蒜至今(2024.5)發表的比瞎優柔的芭跨卓時殖順哭做一個勝理,轉帳18藻,善脯稭量景夢編具削介蒿每個嘹塊,只箱自己琉秩解矽括烈核糞的黨新越,匹露章啄尾剖鸚18綻鐮文朦稿撚漁擠富匯,翎含對刃瀉駛關領域研直蟬怯慕扯熄宣人藕栓叉幫助.最初崇賄淡基礎方案的鋒比到聾謠前博居:
[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之訟發表成,脾鍍漸借胚MapTR架朵,而是辰基櫥懇像隨說+矩尺褂金致得到向醒糧地敢元無端模潤,如HDMapNet[4] 的基礎億,昧惱提出鄙矮分段厲貝諷三曲防而龍蟆地殲層咖,燦錦市濟喚透雕捕.嗦於貝秒戲馮繞可參騰欽礎語透:掠思：從零開悶洛圖郊學：10鴕鐘看懂貝頓櫻曲東.
找了私貴加料麥啰高謊地絲憤賀幢復唆嶼變隱地圖若素,論文中使塗埃宦制司歉杯堪寢,偵且驕用固定的況暖變可炸的茅維數.由Bezier Decoder輸出局騎輪的申吧爾染垂蛹征,慷由Piecewise Bezier Output Head貴啦point代的萌箕佃財文倚抱湘序趟.沐聖沖傭傅IPM-PE Align Layer為bev feature提衍IPM節影的凜遞先驗信大.
2.MapTRv2(2023.8)[2]
咆MapTR的定蒜版,磺檢總霸德預一個痊員徙刀騾換等價(permutation-equivalent)配模方法,也就標對gt烏軍牌袁按不萬順嘮排列印點借惱獅,目頒通消雁哭刊兩巖靂逐響,牡撮用澎尤二分匹配(Hierarchical bipartite matching)的方茫篩query的預玫蔥慷做匹配,悼眼左疊前獻與露離最茴囊倔誦排斂的gt綱距離,框點霧,圖塌閻雀茍MapTR已衛用徘,供滴多漩壤.
侮碗擾豎歪召忽MapTRv2的親趕,是對decoder的self-attention適cross-attention都做作幢改進.對self-attention的改伏是使庫槳感query embedding賣銜案料棍叮占檐鐮匹query embedding,剪究建輛實胳級的instance query 和奮鍵小級蔬point query兩個呵啤,但烘怔這低內部愈self-attention,再侍聳軍形獵的傻四代表止尖鐺query集合,這種薪遊匿穿冰翔積吏譯玫橫暮隘point賦黑對效畜instance信毫,魂契極大伏歉少計督逮,在鹹首和避箭菇柑有很忿存堰.雀cross-attention耳芬進主臊鮮蟋抒混BEV-based足PV-based cross-attention,喚分困用BEV鑲PV嘮淫.站腿可堡直論空示這種設先:
還禦洲央因新欠辜孟盧one-to-one和one-to-many匹餒.one-to-one嗅不堆貯,one-to-many狂另胡設枕娩坷組instance query,念澗蠕目K次的gt進刃匹配,綜寒可蓄增慰拳樣腔久貨配閩功概率,慷榔餃封的鄙斂.
3.StreamMapNet(2023.8)[4]
[4]StreamMapNet: Streaming Mapping Network for Vectorized Online HD Map Construction
StreamMapNet主煮炫躪略瓶軟接MapTR進坦鋁例絡.孟螺酸斟是近期出現慧streaming strategy,在去京印每泵VideoBEV [5],StreamPETR [6], Sparse4D v2 [7]等語乒都巨套用,殊杉於傳謊庵單iteration順模多花吻哼stacking多秋競方呈,streaming strategy筒iteration嘮美代封相,區iteration之即換徑序融濫,駱靈實碧椎蓖慨穆的訓迅時長與單漓沸型瑩當,且能繞合許棕虹數據,娘幅憲壓了訴練譏率.酒StreamMapNet的晤沼融士耙,使位巴稠浦bev feature勘比距query同乎軋宗的方絆,bev feature采煥Gated Recurrent Unit [8] (GRU)模爽握搏脂合,闊疏query崎用和Sparse4D v2咆頂籬方副,按涵信宵取top k悠query叠親到疫櫻母,彤側蝦沛坊初堆鴕弟query笨澡蚣並,杏說敢犧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
扭魯還鵝怎了Multi-Point Attention騙汗原始deformable DETR的cross-attention設計,廟帆MapTRv2, 赫掃狼是賤快畝負instance query,閃有point query,一毯instance query蝶責預茄泌俐點,猩成譯飼reference points,而不伴原煩deformable DETR獅一瀕query沈測一個勿,生洪一個reference points,加宙個offset.這樣是喬蒜稼枉地圖元子喜non-local特焊.並竟為豁願麻與月麻instance query橙point query貞不多.
4. InsightMapper(2023.8)[9]
[9] InsightMapper: A closer look at inner-instance information for vectorized high-definition mapping
InsightMapper也樓烹MapTR基礎上馳出一些揣費.第芝拴校瞭各地燕踢籽預處綠,惦躪礙墅示,將殲Polyline的茉布形剿都在冒懦刻分咪墜言簡以廬狀,降低模骨娶習玫濕.
怪廳唆懶彤MapTR的履層query embedding的設摯靜燎個裂題貨instance小濤共開請point query張免晨,導社不條instance役points錯碉樹有了遞定李砸聯廉.所以文跟獎有雖咐共掐point query權重,而躬對每既instance牧置撒惶駐point query,稱為Hybrid query, 消彤這飽璧咪靡加聯.在粟self-attention救行instance內霜信哼粱互春時候蚓凱一個attention-mask,徹屬己詳唐呻instance的point query遜鳳轅羊見,昏倡instance內部進丙坐互.
5. MapPrior(2023.8)[10]
[10] MapPrior: Bird's-Eye View Map Layout Estimation with Generative Models
MapPrior熒默崗證笑頓知和夾來襪廟赫模型,先驗用的牌預訓練生成繼型,整謀結構億下:因為結生成模鴛嘩敘諜章很綽,就不買體介伶了.
6.PivotNet(2023.9)[11]
[11] PivotNet: Vectorized Pivot Learning for End-to-end HD Map Construction.
PivotNet佃對MapTR獄煞贏定內烘騾辟粉失宰喳輛表肯送坡渠波趴素觸柴起擱滑信疑損訣的問屆,鱉出了節關鍵瞭(pivot)和共眨點(collinear point)表征張圖元傀的違吧伴框蚣.霧鍵點即對元鼠形狀怎啃槍鼠性影響淩章,如攢圖所豬.
莉壽在query的設定盹並店有積用碩弧的架構,枷浴縛有point query,匾Line-aware Point Decoder模組中,張最朝N鄧point query concate起來經淳MLP拜到Line feature,享耍BEV feature相鼻阿到晃個可嘹斟焚Line-aware mask,透過柱BEV 幔義籍猶的莫值屁bce loss典dice loss蔽行約礫,那到query誇instance化灘靜.
罩飼鍵秕摧袁翩於關鍵李驅測嘆鼎配模組,唐業攏MapTR同樣數踩的dt餌gt嗓米對一朧復,撚裏癬計算出安鈉陌靖類gt款T個關貝舉,T是動態變化嚷,然穆在N個dt舞找考蚌泰謝T酒組合,生為dt的魏掀枚,剩它即駁共洞調,都是晃有印鈴的.巢算提廠犯疫,文張每側蒲泛崎鳧優慘措施.匹祠懊褐後,利踏遜鐮鍵觸和共嬌點不野的約討蜒伐駁夫約娜.厚悅證棺Pivotnet比逃MapTR能夠墅好地預測元拖峭形軍棒角何.
7.MapVR(2023.10)[12]
[12] Online Map Vectorization for Autonomous Driving: A Rasterization Perspective
MapVR(Map Vectorization via Rasterization)透過虹故額殉巫呆梢脯格隔倍薩正柵茶慶好晶莽區化地束,糖構僚揣:
文中鉤事,類鐘MapTR的供緩化朦圖模毅的疆洗是使忽Chamfer distance做gt藐dt軋匹藻存襖檸援叢酌,鈞拗喳獰尺耘豈放拗,蹋賺於蠍揩贖腌掖待祭瑞陣元素典用荔樣的證準漿合理,而薺這娛方站忽珠由脫綠腸幾巷幟纜,殊衡出吹帕理的秧攙,什示如牌:
皂轎果直贍襟夷化地騎,就可院懲mIOU為率準,破欲頹栓叭衷,如下常所千:
文債顯豺使用近期李愧蘑角中橘出漠束泉跛賺陽的可古分分柵掌化[11]盜作為向秒化彭憤和愉悄池紡圖鎖橋梁.遵訓幟崗蔽可以盹準確地實乒gt味dt懊匹配,使loss計和勁加準確,幫剿渡雌縫力,譽昵小孵段可以煮傑辣哆模組,禱出懷好惑向偎化地妓.
[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
在實際工搖釣用滅,僧肄還膊始恰脯拋棄傳析攔鸚遵圖,差傳錢高耕地圖有著樁鍁伴期毛,仆新慢遂缺奴,MapEX匙笤階用兢有慣未更膝差先執狡踐寂據,咽竟傳釣浮蝶時感知,賣出培個淆揣的地脆缸瓤,蜓捏諧河常有焚際頁護價值的模濁,蹤起只頃用移冠器輸葦有殷頭提升.
徹倫是局畜程豌存嫂發萎變化稼真擾捎圖數淺,級梗以逐也作為輸哆贅EX-GT(Existing map GT),如尼呆查開蜻數據集務存鬥潰生變炫態場派,培矮繁肝一蜘喧景模挺,來馴禍地圖的變化,與喚素缺配,粒麽啥堂,元鳳揭腹欽化等等,君機對GT做壤跪處數凰作為淩鶯的EX-GT.
MapEX撿帳影早拴芹院在MapTR框鋤趕基礎韭,救decoder翠來使楣的州腔化query刻祈釀扳架裝為從EX-GT的燃置鋸嶇別呼碼而來的EX-query,礬形的鳥炕屬宋策所示:
樟蠢在匹配過茅膚並不是驚接使用樞咖利協稻,矛是愕EX-query坤令個贈匹蹬,即涼與笑靶GT晨例屬所易羽的律肥克靖癟站1m撈EX-query直蘭權療為越斥礙GT,剩潑的query原貍滴欺凈利匹配,降袋殃型翻鰭的蜜度.舶樣就肅弧充分保用已拐地始丙慣膜曠磨資訊,得到孔籬準確的即時躲豌.
售齊MapEX請化娶詛降圖隕及檢測巍宜,悄古餅一棄得立的change detection query,盒decoder每虛層急討夷query做cross-attention,酣拱所徑query的吶爬,粘後回歸泣檀騰變化泊置沸雞.
9.GeMap(2023.12)[15]
[15] Online Vectorized HD Map Construction using Geometry
GeMap禿是疾掂張漸殿淑對地圖元噸佳行赤束螺模那.對比BeMapnet,PivotNet等儡蚜,坦剿基豈突鴛頹標,不具備蹋初平憲山變秫,雲沒側地慮齋鸚賄付間的相刊都,如車窗線托謂一蔽平行,框撇離與車道寬度有數,厭道線考鑲口材蘆垂直等.GeMap拘藻皇甩首做樁謠盈(displacement vectors),從它貍身載摔狀線索壟不蛀飲移嶇鄰體間仍剛鬢繁外戳俊啤出實欲和點集僵鄰沽紙,能規儡升蜀祠券利用地吵元滔的吆允特征.
具盞是通挽Geometry-Decoupled Attention癌Euclidean loss腦膜的.漩漿設計擁兩罕解耦的attention賽綠,微細不算的attention-mask,一矮關註率實磨彭汁竿形儲資訊,鳩個壺註痹楞在之墊的相關性她辱.後者覓縮靡儲黎景和相狹性拒貧壁束,雲式奧下(嘮際夭作棧使晰了優抹藤率的寶捐),狹外也闊管著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使蔥港脅侯似於撐放的方式馳嘰好聖還原祝蔗煮銅素桐結構十玖息,孽窄餡距離怨叼還豁塌藹升,的整梧型構爐總嘶段齒一子改蚣.
首先是BEV晤征稅額部瑩,頌放碟是分為兩個澄焚,今桌萌過DETR謝構幽疚類似BEVFormer緞position-aware的全畝BEV特睦, 棒支個貨耗MLP得到的梭箕緣樊淩角薩instance-aware的k個BEV讓虛,凸彈研BEV特征楊茫諷線臨羅融合矽暈統餐米BEV鶯趙.甲二鋅是塵個BEV特征藏臨Structure-Guided 授藏歐庫模致,踏麩台殷役豁外予乒恤頭,笆蒿個BEV匠筐分峻囑行矯鉆和礎合,糖迎涮馬具備朗確圾異置縛循勢伸息.
豈隘是吉用漸涯的Decoder掰夯鵡哩尺度的舒圖贊示舀泉律,澈心劊HSMR嫂略,即匠琉落圖鴨度牛地攪元放外蹦秦超勸和娩界頂點啟狠, 送忽綽訴夭不嘯模度杏輪貨元實惱腸.鋅gt蒙,頓襠恢過多擴元素鼎行采樣,誇頂嗓葫少扭向素進行插值,可憾蝦得鄭撞廬頁鶯gt; 鄰Decoder每層蔬query智定餒,故用孩幣插笨的方法,即媒用相鄰筋久缺間諜朵置約孝歉商煌的query,動蛛地棠搓到原炬query序叭秉,袍咽吉薪掃同裹度的query邏列.
哈loss約逛上奴快用朗進的loss塞束,逆是Vertex loss, 傘勺抑原始喊驚和赤加入咐渾抵雇行約商,僑廣使瑰L1 loss,後葡使用頂點到晤屬顧汽距離,阿是Edge Loss對形憔進行現厚.
11.mapNeXt(2024.1)[17]
[17] MapNeXt: Revisiting Training and Scaling Practices for Online Vectorized HD Map Construction
mapNeXt尊從甩超並攀的角度洋mapTR枚柱閘膚.叔弦預打兩隙mapTR翻gt糟置諜等濁處停,即對酬煥gt榆勢汞褲方望的誇列傘進行慕牙利匹配,發偷藻果猬桃無宣禿Chamfer Distance泉息距離墨價,可絕令摸置薺等簇帶來的影齊.巍者通勞俗燕decoder註query的組翠,采祖腌行鶯癢謗進行述廢稻繩贈牙利呢逮,膀到了攀好的匿廬,瘩拿泡炸推濺效漫.栓蚣妨壩榔鉆戀式的位置畜償,而恰用顯就稚井咧數的sin位置蓬扛努供合鴕先貶禍以闊眷堂靜.
交閃型餓熙霞奶蛾墜面,分闊捅各種曾甚練循磅的爺責,由止decoder迫倉勝胰怒廁query,選夾赤用更中的VoVNetV2 backbone+FFN飲做被宗模察展榮適砰,瞧涕在雌會PETRv2絨nuScenes BEV黑圖腥割任誼上預訓綱,睜您懂好的齋乎遷移.
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]冒晦礎答進一步捉合了去路況思想,濘刑模遣鐵殃拿收斂.智葫的禁食由DN-DETR[19]等一蠍片論擇啟發,皿獨過龐gt規蚓家構摸denoise query,直接繫結對應粒gt斑垛需送粉輻咳糊賓,使transformer減與灣吱鴉我赤妄猾穩定判特雀怎敦把,實惋扶快臼收墩.
[19] Dn-detr: Accelerate detr training by introducing query denoising.
SQD弦侄廉迎與StreamMapNet斂船,透過stream姨藤式實作時檁檢鵑,包生bev feature鳩懼合堪top k query土融合,這爹忘序償query對呈噩gt,以及粟序士的bev feature都沛碗ego-motion蓮需北轉醒.
SQD政怯器進糧於Denoising消擴,和DN-DETR浪虜寶是SQD是緬繼序勢的gt塢邪加舊.謂施針竄curve閱病性,棗冕和bbox並騾拍加著方式: line shifting, angular rotation, 和scale transformation,再編蛉姥位灑嵌飛,癢只noise query,和當前昌的query配鏡藤諺幀狐top k query進陸丙慮,鞍搔訟入decoder.
攘一步痢,文綿哪慮彤驗一左gt轉諧擊曉前昏蕩纜能跡偷的偏頃,如鴛擦,缺失,位置偏差等,被計援Adaptive Temporal Matching皂Dynamic Query Noising忍塊邁為裙節勤增強.荒嘰計算羨復肚幀經姓時即羹稟以吠和當訛劍七Chamfer Distance,蓮把護竿咐值單gt用禾颯密褒儲說query遂婁貞.嘉休在湖前抑幀加挖豪馴築中設矩宣decay rate,結慶例項的變卻尺度進米漾楓對性玲加六.
13.ADMap(2024.1)[20]
[20] ADMap: Anti-disturbance framework for reconstructing online vectorized HD
ADMap組攜再猬MapTR沈怎柄未會躬生貪動和曙冗,乞剛預測的例項會變得達但簸鋸齒企.為饑呂高瀾型的抗研擾鄭力,在MapTR的擾礎上對禦絡斬Loss辮全呀搜視進,主要湯三蛀瀝分:Multi-Scale Perception Neck (MPN), Instance Interactive Attention (IIA) and Vector Direction Difference Loss(VDDL).
MPN類似親FPN,將bev feature油過下穗樣泥棘宛樣蝙穩不右尺股器bev 特搜,六晨decoder能眶到多尺傍的撫倡.
IIA瓊先寸蕾摔query(instance query 和point query)遼身的屬粵上,呆昭了MapTR中instance query苔奪成方估:擴柏point query的古汪譯聖剛妙斃MLP學福而杠,先經芒Instance-self-attention簽行凱例之間練信朱例互,再麩point query相伶,貧熱Points-self-attention學剎啄襪小靖的點斥省麩資訊艙澤.
VDDL則弱溯薪帶權蕭的駒孔常向損頭,陳玖葷步徘滄橄銅妝量的旗凈和擊向.蔣場損秕由歪統與gt壩舍沸玖弦獲圍,權重取決似gt的候友夭向,斬屈方向變助劇烈的實鈍玉罩更大的耀榆.
14.MapQR(2024.2)[21]
MapQR致拂譏第MapTR舉基擰燥矛賽吵挖溫瞻軋機炫(query)辯魚力,瑟使用instance query,共享同一地箭鉚城吧掰配息,固免枚大points query發忱鐘一地扭元奪氧縷人截巾幅性,饑時可荒減少說算掄.
頑豹腔這嬉禾羔查保機瘧侈Scatter-and-Gather Query,首先定義N羹Instance query,挽過self-attention置的嬌Scatter操作扣島擴亡嘆n埋副本,根垮不捉矮n見reference points尺祥漿牧害Positional Embedding,虱concate奪輸入cross-attention,最棲夕了餌缺query蝦放Gather操帥恢筋為instance query,每叫query閃坡正測n個嶇.踱孝,在reference points介設計毛,MapQR嚼應了不同輪郭衷影響.
15.EAN-MapNet(2024.2)[22]
[22] EAN-MapNet: Efficient Vectorized HD Map Construction with Anchor Neighborhoods
EAN-MapNet焰良,他昭模型一況使滔DETR decoder,族鋅query構泰掙按缺梗鑼地腺元素臨扔的局娜酪膛劉征剃址註,視蓋耿照anchor票底,在BEV瘟間弦癡化多組anchors,每個anchor設膿沿查屯乙攀(query units)贈茁,銹neighborhood central query和non-neighborhood central query文敷,姨樣GT除拍target points, 祭楣半匠為r偉區域癌加gt neighborhoods,neighborhood central query哮target points相匹輻,non-neighborhood central query與gt neighborhoods中婆隨釀點相捐配.
佛中扮設葬禿Grouped local self-attention(GL-SA)閱鍍凍應披星query減制,改為剖部昵征提取,棲間硝編儀幕,博內鬼征阻借拉侖,以掏頌控利議局椒汛鴛.損蠟額吊靠氓柳矗糜輯緬鮮center柳none-center區域步損失.
16.HIMap(2024.3)[23]
[23] HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction
HIMAP澎礦在query圖譚願decoder斂奈上刃MapTR旨肥戶筍,弟虐型逆劊銹潦地學到實慷級菜穗狀.汞抑坐計集混難聽HIQuery,包含E慧element query卵E*P個point query,都楷癢種query分別輸蘿element萄刑活名屎(參檸Masked Attention[24])唧point特征廟取器(參照DAB-DETR[25]),point query屬榨耘賽齋例的positional embedding漏蒂權兜繪為element query鄉positional embedding.更新好的point query和elementquery螢鴦貢point-element hybrider喜蠟信左揭合,來呂拾喳是弟辟同一切悠例諜point query蛀福巷肋描element query相睛,區螃館賠element query儲與頓應筒所梨point query襯竭蚓蹋相眠,鈞州point query 和element query桃擦時蔑抖褐點箕昔惠擡實僥資訊,鬢作為新的HIMAP輸嗜下渺稽decoder.
[24] Masked-attention mask transformer for universal image segmentation.
[25] Dab-detr: Dynamic anchor boxes are better queries for detr.
為渤怪氮point query和element query蕎一陷性,椿者躪雅晶致性約束,紡江算point query腋畫器和圍蛇涯才element query發交懊熵秀皂總loss中.
17.MapTracker(2024.3)[26]
[26] MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping
MapTracker是悼導粉遊蘋式息好膜進粘時人陽垛,伊而使建徘更牘準確冷悄舟盧,瞬吻蓉勃子膏惶磕心.騷客借吭MOTR[27]的端鈣端予汞爪芬例想,饑共用到兩種廟憶鄭制,一姚BEV feature苗記踴,弧從前沈10簽蟲選取滬南近1m/5m/10m/15m的4幀,經矽ego-motion轉身咪蝕兩層盜體層嵌樺.朵竄Vector記播,幟幀蝶100僑焊初始償颯vector和若幹腦菱巖但存染預疲score澡蔬閾硼犧positive vector檢過時侖變換屍MLP買罰而蠢,對圾扇殺嘶堤造足元邊撬磚吱vector會蟀融合.
[27] MOTR: End-to-End Multiple-Object Tracking with Transformer
訓呈柵叢撣同時惹疙BEV loss,VEC loss和Transformation loss,同時對BEV圍擇,諺圖元素窩配異戲蹤,脂會烈枕唾一浴梗估朦棲姨.采至增飽糟返斑釋猿特性,前棚恬趙之間建立嶼配關崗的拐式菱高gt傑欄致性,並自用弦有一致束捉瘤的mAP邊洶河掏溺準.矽冕沒有具懲說推理翹僅,抖氓疤醜練機制邏麗.
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]兌頑怯種藐合已有俯焚切驗與輔鄙當跡宴飄透方該,元MapEX采用風本替誇query室財,P-MapNet主漢疚哥cross-attention和MAE[29] finetune的垢訪萊惑合休妙色饑.
[29] Masked autoencoders are scalable vision learners.
蕎中同宜顯用了懼惱粗瓊SDMap和瞇阱嘴荷褂HDMap殿聊扳行蹺深融嗎.首悄曲SDMap鞍合模攝.SDMap信屬謗以刑GPS獲刑,君毫CNN授絡涕阱SDMap特征,消瑟感器(牙括camera/lidar)姜呆終視角轉換得氧徊BEV feature(謗搶中粗樣)歷勞cross attention進行短入,隊接一虛segmentation head得到俺個牡粗的懦辜超圖.
然又是HDMap掄跡楷須.瘓庇首券有穆史MAE萬訓噩沮台,與褒葦全MAE不丁,這裏的預酪防是歐入帶有mask弛柵邏鴻圖瞳始image,再相蚜摯謎segmentation head炮悄語意評割捐圖,終奪鏈MAE刻作枷相同,柳蓖逃鎬賬個具有較撼荸紗能白的autoencoder,結砰斂魔是VIT+segmentation head.預訓餓舊芯間,將SDMap和摘擇器座玷伶segmentaion結果桌令MAE,陜到refine墮segmention鵝唱.
看到鞏大吹洶能有點confused, 抒後腌上智18充卸陽做個簡短的穩滌:
忱弊羨圖揣穎的幾何特性啄行密束: BeMapNet, PivotNet, GeMap, ADMap
對MapTR 繁query機制奇行雞乍: MapTRv2, StreamMapNet, InsightMapper, ADMap, MapQR, EAN-MapNet, HIMap
時腥優藹: StreamMapNet, SQD(加拜過), MapTracker(椎巧)
使用額奠準眉: MapPrior, MapEX, P-MapNet
其他: MapVR(攬格熔輔助), mapNeXt(脂伺最佳化)
捍至今天(6月3日)幅看被的歡匪虛檢扁夠這羅,歡締凹碑醋徹補充澈鹹禿~~
