@camstack/addon-pipeline 1.2.294 → 1.2.296

package/python/postprocessors/testdata/ssdlite_mobiledet_outputs.json

@@ -0,0 +1 @@
+{"_provenance":"Raw output tensors of google-coral/test_data ssdlite_mobiledet_coco_qat_postprocess.tflite (sha256 32c486140391eb4dc43fca7113ad392be632dc5366687f2731f73d740678693f; the CPU build of the same QAT graph as ssdlite_mobiledet_coco_qat_postprocess_edgetpu.tflite, whose four outputs have the identical names, shapes and dtypes), captured 2026-09-26 on ai-edge-litert 2.1.3 through inference_pool.py's edgetpu load/preprocess/predict path with a CPU interpreter, letterbox 320. Frames are COCO val2017 images; only the model's numeric outputs are stored here, never the pixels.","frames":{"coco_val2017_000000039769_two_cats":{"orig":[640,480],"scale":0.5,"pad":[0,40],"outputs":{"0":[[[0.15955263376235962,0.534464955329895,0.713003933429718,0.9716978073120117],[0.2043571174144745,0.02350570261478424,0.8648291826248169,0.5011452436447144],[0.10257476568222046,0.014032959938049316,0.9505413174629211,0.996590256690979],[0.22124464809894562,0.04488908499479294,0.3162550926208496,0.29261064529418945],[0.03658798336982727,0.012482762336730957,0.40508130192756653,0.987517237663269],[0.1555221676826477,0.6529091000556946,0.8028101325035095,1.006462574005127],[0.24686561524868011,0.5220372080802917,0.4193393588066101,0.5692116618156433],[0.2072891891002655,0.1583276093006134,0.3572945296764374,0.4583382308483124],[0.23659078776836395,0.5290685892105103,0.42224830389022827,0.5870484113693237],[0.5743775367736816,0.23019926249980927,0.6594175100326538,0.34779107570648193],[0.3052724301815033,0.24310186505317688,0.9099631309509277,0.9973288774490356],[0.11994627118110657,0.3792511224746704,0.5946398973464966,0.9894982576370239],[0.2084560990333557,0.2897113263607025,0.3498774766921997,0.4769533574581146],[0.048862308263778687,-0.024405688047409058,0.8030894994735718,0.6076090335845947],[0.47406071424484253,0.039728641510009766,0.8791598677635193,0.9857875108718872],[0.20506274700164795,0.006122678518295288,0.3345203399658203,0.26888036727905273],[0.21482816338539124,0.2670758366584778,0.4976709187030792,0.48709094524383545],[0.17930985987186432,0.5188078880310059,0.6031879782676697,0.87400221824646],[0.5569137930870056,0.8219212889671326,0.6033983826637268,0.91482013463974],[0.5594297051429749,0.06481415033340454,0.6947370171546936,0.3435206711292267],[0.2410658895969391,0.526638925075531,0.3487098515033722,0.5821121335029602],[0.24299399554729462,0.8754947781562805,0.3305767774581909,0.9966028332710266],[0.09360367059707642,-0.027666032314300537,0.4855625033378601,0.7005844712257385],[0.21883206069469452,0.37879708409309387,0.3296947479248047,0.46383705735206604],[0.2706087529659271,0.8736841082572937,0.34845462441444397,0.9984135031700134],[0.23773013055324554,0.5050985217094421,0.5164366364479065,0.5824022889137268],[0.8736730217933655,0.004488527774810791,1.0054889917373657,0.9902228713035583],[0.23501813411712646,0.1322900503873825,0.3002501130104065,0.2705683708190918],[0.2720530033111572,0.9462981820106506,0.35584938526153564,0.9963359236717224],[0.247519388794899,0.944722592830658,0.3231050968170166,0.9993845820426941],[0.38305434584617615,0.378726065158844,0.47873106598854065,0.49631792306900024],[0.5535222291946411,0.7801342010498047,0.7023705244064331,0.8831243515014648],[0.7894991040229797,0.03270852565765381,0.9892374873161316,0.9928076267242432],[0.20900268852710724,0.05372738838195801,0.29099857807159424,0.3962726294994354],[0.2068709284067154,0.006370965391397476,0.3372364044189453,0.0244777649641037],[0.23342448472976685,-0.01042163372039795,0.3790746331214905,0.2895890474319458],[0.22075217962265015,0.22743815183639526,0.8087112307548523,0.824122965335846],[0.5976130962371826,0.05392265319824219,0.9607176780700684,0.9731613397598267],[0.11443797498941422,0.7428056597709656,0.24389559030532837,0.9905272126197815],[0.22740966081619263,0.4848347008228302,0.4475909471511841,0.5901657938957214],[0.21618488430976868,0.5098341703414917,0.44631537795066833,0.6151652336120605],[0.8084112405776978,0.0015302002429962158,0.9273114204406738,0.9038747549057007],[-0.003454655408859253,0.01914122700691223,0.29720693826675415,0.9142001867294312],[0.23717385530471802,0.4077231287956238,0.6862777471542358,1.043309211730957],[0.40243661403656006,0.4213453233242035,0.46671468019485474,0.49919119477272034],[0.038508474826812744,0.13357895612716675,0.5688574910163879,0.814859926700592],[0.2079186886548996,0.4072537422180176,0.25239354372024536,0.45305848121643066],[0.18039143085479736,0.7189949750900269,0.346691370010376,1.0018376111984253],[0.17740964889526367,0.8574067950248718,0.3975909352302551,0.9967599511146545],[0.30818599462509155,0.000822313129901886,0.6126490831375122,0.1720949113368988],[0.06299805641174316,0.032074540853500366,0.2399788498878479,0.8108158111572266],[0.13213692605495453,0.5332436561584473,0.41890597343444824,0.9667563438415527],[0.3001849353313446,0.8360094428062439,0.5748157501220703,1.0098237991333008],[0.46447721123695374,0.7584561109542847,0.6980230808258057,0.8998774290084839],[0.22825893759727478,0.5397037863731384,0.7288434505462646,0.8317379355430603],[0.21569259464740753,0.8997492790222168,0.2564050257205963,0.9679288864135742],[0.25803816318511963,0.4330609440803528,0.3111131191253662,0.4992610216140747],[0.22732794284820557,0.3999912738800049,0.4643397331237793,0.5687592029571533],[0.23802617192268372,0.5667713284492493,0.42239078879356384,0.8623947501182556],[0.34699746966362,0.2612840533256531,0.488418847322464,0.5053830146789551],[0.5587438941001892,0.22770456969738007,0.6354513764381409,0.34881260991096497],[0.5473014116287231,0.6006264090538025,0.6144839525222778,0.6729443669319153],[0.5691072940826416,0.5631829500198364,0.6691073179244995,0.6897634267807007],[0.197118878364563,0.5749090313911438,0.5393403768539429,1.0084216594696045],[0.2407025694847107,0.4220343232154846,0.41971439123153687,0.648799479007721],[0.43011459708213806,0.42714160680770874,0.49631467461586,0.4729463458061218],[0.5331307053565979,0.7939615845680237,0.61097651720047,0.8855019211769104],[0.2184237539768219,0.7220790386199951,0.6754136085510254,1.0115272998809814],[0.23764437437057495,0.8479610085487366,0.8712865710258484,0.9984578490257263],[0.12035168707370758,0.8447889089584351,0.35048216581344604,0.9947942495346069],[0.21286840736865997,0.9534194469451904,0.26364877820014954,0.9906877279281616],[0.23123027384281158,0.04802142083644867,0.31140387058258057,0.12702254951000214],[0.5154188871383667,0.7573367357254028,0.6169029474258423,0.8059217929840088],[0.697390079498291,0.05803816393017769,0.8673416376113892,0.11111311614513397],[0.22868293523788452,0.3173567056655884,0.32868295907974243,0.45768725872039795],[0.24360230565071106,0.01714465767145157,0.5480654239654541,0.2307719588279724],[0.2897666394710541,0.5481002926826477,0.4560665786266327,0.839400589466095],[0.37094348669052124,0.7636645436286926,0.6415581107139587,0.9050858616828918],[0.46699267625808716,0.5482616424560547,0.6871740221977234,0.7246556282043457],[0.27251920104026794,0.03307163715362549,0.827480673789978,0.32943978905677795],[0.03661632537841797,0.2884189486503601,0.7459261417388916,1.0189385414123535],[0.05953405797481537,0.8303914070129395,0.2862992286682129,0.9966913461685181],[0.17441785335540771,0.2090485841035843,0.27974894642829895,0.5367846488952637],[0.18664157390594482,0.5720193386077881,0.35294151306152344,0.919648289680481],[0.19373293220996857,-0.003133825957775116,0.531268835067749,0.1656341254711151],[0.27887362241744995,0.9335945248603821,0.658628523349762,0.9955716729164124],[0.32571566104888916,0.0015931539237499237,0.6534517407417297,0.07340744137763977],[0.4985097050666809,0.058443158864974976,0.6723241209983826,0.32905781269073486],[0.5864514112472534,0.15539874136447906,0.6753339767456055,0.29165494441986084],[0.5612818598747253,0.6305587291717529,0.6754595637321472,0.719441294670105],[0.513027012348175,0.5583153367042542,0.7057222723960876,0.8958513140678406],[0.004697104915976524,-0.005696093663573265,0.040883369743824005,0.03212530165910721],[0.20103901624679565,0.33512550592422485,0.35104435682296753,0.5940406918525696],[0.20186400413513184,0.4540380537509918,0.37313663959503174,0.716795802116394],[0.22820225358009338,0.3030018210411072,0.40721407532691956,0.5136640667915344],[0.43144962191581726,0.2230268269777298,0.6185503005981445,0.4936414361000061],[0.48286062479019165,0.7618263959884644,0.6039686799049377,0.8043785095214844],[0.5385960936546326,0.6290205121040344,0.6187697052955627,0.7091941237449646],[0.4680514335632324,0.7121118903160095,0.7194494009017944,0.871222198009491],[0.6098172664642334,-0.0005137994885444641,0.7401827573776245,0.04462110251188278]]],"1":[[16.0,16.0,64.0,74.0,62.0,16.0,74.0,16.0,43.0,73.0,64.0,16.0,16.0,62.0,64.0,83.0,16.0,16.0,84.0,16.0,74.0,74.0,62.0,16.0,74.0,74.0,64.0,83.0,74.0,74.0,74.0,84.0,64.0,74.0,83.0,83.0,16.0,64.0,0.0,74.0,74.0,64.0,62.0,16.0,74.0,62.0,74.0,31.0,74.0,83.0,62.0,16.0,0.0,58.0,16.0,74.0,74.0,16.0,16.0,16.0,73.0,76.0,76.0,16.0,74.0,74.0,84.0,16.0,64.0,0.0,74.0,83.0,84.0,74.0,16.0,83.0,16.0,83.0,39.0,16.0,62.0,0.0,16.0,16.0,83.0,0.0,83.0,16.0,74.0,84.0,30.0,0.0,16.0,74.0,16.0,73.0,76.0,84.0,31.0,83.0]],"2":[[0.76953125,0.76953125,0.5234375,0.39453125,0.30078125,0.24609375,0.23828125,0.19140625,0.18359375,0.17578125,0.1640625,0.15234375,0.14453125,0.14453125,0.140625,0.1328125,0.1328125,0.125,0.11328125,0.11328125,0.109375,0.109375,0.109375,0.10546875,0.10546875,0.10546875,0.10546875,0.1015625,0.1015625,0.09765625,0.09765625,0.09765625,0.09765625,0.08984375,0.08984375,0.08984375,0.0859375,0.0859375,0.08203125,0.08203125,0.08203125,0.08203125,0.08203125,0.08203125,0.078125,0.078125,0.07421875,0.07421875,0.07421875,0.07421875,0.07421875,0.07421875,0.0703125,0.0703125,0.0703125,0.06640625,0.06640625,0.06640625,0.06640625,0.06640625,0.06640625,0.06640625,0.06640625,0.06640625,0.0625,0.0625,0.0625,0.0625,0.0625,0.05859375,0.05859375,0.05859375,0.05859375,0.05859375,0.0546875,0.0546875,0.0546875,0.0546875,0.0546875,0.0546875,0.0546875,0.05078125,0.05078125,0.05078125,0.05078125,0.05078125,0.05078125,0.05078125,0.05078125,0.05078125,0.05078125,0.046875,0.046875,0.046875,0.046875,0.046875,0.046875,0.046875,0.046875,0.046875]],"3":[100.0]}},"coco_val2017_000000252219_street":{"orig":[640,428],"scale":0.5,"pad":[0,53],"outputs":{"0":[[[0.4162486791610718,0.024550773203372955,0.7795857191085815,0.1983652412891388],[0.4350929260253906,0.5026833415031433,0.7440744638442993,0.6160659193992615],[0.4274504780769348,0.8110100030899048,0.760050356388092,0.9848244190216064],[0.5311018228530884,0.3237488269805908,0.7480632066726685,0.5054173469543457],[0.6483170986175537,0.6389470100402832,0.7928438186645508,0.7346237897872925],[0.5338170528411865,0.3173767626285553,0.6368073225021362,0.3987402021884918],[0.4912804067134857,0.9553788900375366,0.6911295056343079,1.0005137920379639],[0.30877313017845154,0.8685252070426941,0.41176334023475647,1.0109384059906006],[0.28655675053596497,0.8688049912452698,0.5342783331871033,1.0041122436523438],[0.5341765880584717,0.30358779430389404,0.6143501996994019,0.3802952766418457],[0.24573861062526703,0.535965085029602,0.3307785391807556,0.599303126335144],[0.4633682072162628,0.9736822247505188,0.6661832332611084,0.9969421029090881],[0.5260692834854126,0.38162389397621155,0.6489746570587158,0.48016151785850525],[0.7162842154502869,0.5424903035163879,0.7484474778175354,0.6030901074409485],[0.3242916762828827,0.9250457882881165,0.40213754773139954,0.9984369874000549],[0.5646662712097168,0.32681071758270264,0.6617629528045654,0.4069843292236328],[0.43745797872543335,0.8556292653083801,0.506649374961853,0.8928975462913513],[0.5453707575798035,0.9623616933822632,0.6898974776268005,0.9964772462844849],[0.3656116724014282,0.9378393292427063,0.6885550022125244,0.9913268685340881],[0.6316292881965637,0.35618144273757935,0.7316293120384216,0.47908681631088257],[0.299285888671875,0.9094834923744202,0.39638257026672363,1.0051602125167847],[0.5499159693717957,0.13722729682922363,0.5882986187934875,0.20540684461593628],[0.5260947346687317,0.3105463683605194,0.6739063858985901,0.4977883994579315],[0.5227016806602478,0.9099226593971252,0.6287714838981628,0.9592286944389343],[0.5940675735473633,0.3829779624938965,0.6574056148529053,0.46554887294769287],[0.4623275697231293,0.8369593620300293,0.4979846775531769,0.8631285429000854],[0.5491682291030884,0.47726282477378845,0.62587571144104,0.5139861106872559],[0.2056179791688919,0.4550957679748535,0.2488015741109848,0.5022701621055603],[0.21111512184143066,0.5228007435798645,0.24625077843666077,0.552243173122406],[0.5356869101524353,0.49689340591430664,0.6069472432136536,0.5310089588165283],[0.5498236417770386,0.4632295072078705,0.7031227350234985,0.511814534664154],[0.46009230613708496,0.7904802560806274,0.7690738439559937,0.9199378490447998],[0.5481249690055847,0.4862275719642639,0.7283921837806702,0.5313624739646912],[0.5189509391784668,0.02130730450153351,0.7701113224029541,0.44223350286483765],[0.6050348877906799,0.9128004312515259,0.6596968770027161,0.966663122177124],[0.4300398528575897,0.8400573134422302,0.4641554057598114,0.8629769682884216],[0.3168220520019531,0.9083377122879028,0.6748456954956055,0.9979112148284912],[0.5227682590484619,0.8477926850318909,0.6257585287094116,0.913992702960968],[0.7170714735984802,0.3222942352294922,0.7432406544685364,0.360115647315979],[0.49940651655197144,0.5581434369087219,0.5609056353569031,0.6006955504417419],[0.48628464341163635,0.40381699800491333,0.730383574962616,0.5295159816741943],[0.5414602756500244,0.5834677815437317,0.6011738777160645,0.6092543005943298],[0.7611274719238281,0.0054911356419324875,0.8006578683853149,0.03861616551876068],[0.5396264791488647,0.10698680579662323,0.760373592376709,0.5930132269859314],[0.17262639105319977,0.8333241939544678,0.4315415620803833,1.0020921230316162],[0.48002371191978455,0.8558622598648071,0.5729225873947144,0.9073963165283203],[0.5417873859405518,0.05324568599462509,0.6141053438186646,0.11295924335718155],[0.5150630474090576,0.8205909729003906,0.6275711059570312,0.8500334024429321],[0.5987178683280945,0.4551270604133606,0.7453894019126892,0.5037120580673218],[0.23898977041244507,0.5919873118400574,0.33752739429473877,0.6653785109519958],[0.2905657887458801,0.5573469996452332,0.3343902826309204,0.6073847413063049],[0.41652873158454895,0.8505421876907349,0.4688275158405304,0.8861992359161377],[0.4871012568473816,0.10838092118501663,0.5746840834617615,0.1622435450553894],[0.5822117328643799,0.850226879119873,0.6545296907424927,0.914504885673523],[0.6951959729194641,0.34702178835868835,0.737125813961029,0.4853001534938812],[0.5352706909179688,0.10564260184764862,0.6014707088470459,0.17087455093860626],[0.6321781277656555,0.9626111388206482,0.6927779316902161,0.9962278008460999],[0.7240806818008423,0.4945259094238281,0.7494900226593018,0.5657862424850464],[0.7278625965118408,0.6477597951889038,0.7912006378173828,0.699293851852417],[0.44048047065734863,0.9741995334625244,0.5787588357925415,0.9905321598052979],[0.5262652635574341,0.526103138923645,0.5561445951461792,0.550413966178894],[0.5032247304916382,0.8301358819007874,0.5570874214172363,0.8566934466362],[0.5438975691795349,0.31178784370422363,0.688424289226532,0.37798792123794556],[0.741370677947998,0.0036882534623146057,0.7968438863754272,0.1787216067314148],[0.26776328682899475,0.8388197422027588,0.4030705988407135,1.0153470039367676],[0.4859636723995209,0.8949870467185974,0.5802412629127502,0.956486165523529],[0.6024523973464966,0.49162277579307556,0.7387086153030396,0.5348063707351685],[0.5774096250534058,0.6351754665374756,0.797590970993042,0.762739896774292],[0.2301507294178009,0.5263333320617676,0.3286883533000946,0.5708081722259521],[0.2292805016040802,0.6279017925262451,0.3369244635105133,0.6867420673370361],[0.5606799125671387,0.027291707694530487,0.5996322631835938,0.05806448310613632],[0.5880016088485718,0.5785675644874573,0.6782033443450928,0.6112081408500671],[0.718255341053009,0.01484508067369461,0.7508959174156189,0.13957443833351135],[0.7315396666526794,0.8035227060317993,0.7597091794013977,0.837638258934021],[0.2944031357765198,0.809653639793396,0.453513503074646,0.9861807823181152],[0.6220482587814331,0.33179566264152527,0.7205859422683716,0.4476678669452667],[0.42942866683006287,0.7300208806991577,0.7904558181762695,1.0069555044174194],[0.44813793897628784,0.14109531044960022,0.46520859003067017,0.15604627132415771],[0.5543093681335449,0.5982905626296997,0.6266273260116577,0.6222459077835083],[0.4242232143878937,0.5523462295532227,0.7424439191818237,0.6393190622329712],[0.5669523477554321,0.5963963866233826,0.6992526054382324,0.6285596489906311],[0.21041908860206604,0.4460771381855011,0.34374770522117615,0.683089017868042],[0.5402801036834717,0.27855896949768066,0.7388873100280762,0.4006084203720093],[0.5964339971542358,0.8200300335884094,0.6535660028457642,0.8447017073631287],[0.2541125416755676,0.4373975098133087,0.3120924234390259,0.5199684500694275],[0.5770773887634277,0.04363440349698067,0.6699762344360352,0.12993639707565308],[0.7549723386764526,0.26059067249298096,0.7994471788406372,0.3144533038139343],[0.39701393246650696,0.9599695801734924,0.46035197377204895,0.9885571599006653],[0.5154039859771729,0.14906154572963715,0.5684789419174194,0.20535801351070404],[0.5914602875709534,0.08590969443321228,0.6511738896369934,0.16261713206768036],[0.6257311701774597,0.5529274940490723,0.7242688536643982,0.602965235710144],[0.7689149975776672,0.748873770236969,0.7928703427314758,0.7981798052787781],[0.212331622838974,0.4396534562110901,0.24356108903884888,0.46190768480300903],[0.5010358691215515,0.03401685133576393,0.5607494711875916,0.046919818967580795],[0.5261155366897583,0.33853837847709656,0.5592405796051025,0.390837162733078],[0.43047335743904114,0.00413740798830986,0.7486940622329712,0.07919672131538391],[0.5710752606391907,0.0036823004484176636,0.6921833157539368,0.06252261996269226],[0.6423197388648987,0.6038827300071716,0.8014300465583801,0.7752847075462341],[0.5016130208969116,0.3145033121109009,0.7622942924499512,0.5751845240592957],[0.497750461101532,0.720516562461853,0.561088502407074,0.7368491888046265]]],"1":[[0.0,0.0,0.0,61.0,32.0,26.0,0.0,27.0,27.0,26.0,0.0,0.0,26.0,40.0,27.0,26.0,0.0,30.0,0.0,61.0,27.0,30.0,26.0,30.0,26.0,0.0,26.0,0.0,0.0,30.0,61.0,0.0,61.0,14.0,30.0,0.0,0.0,30.0,36.0,30.0,0.0,30.0,40.0,14.0,27.0,26.0,30.0,30.0,61.0,71.0,0.0,0.0,26.0,30.0,61.0,30.0,32.0,40.0,40.0,0.0,30.0,30.0,26.0,40.0,27.0,30.0,61.0,32.0,0.0,61.0,0.0,61.0,40.0,36.0,27.0,61.0,0.0,76.0,30.0,0.0,61.0,71.0,0.0,30.0,83.0,30.0,36.0,27.0,42.0,30.0,40.0,40.0,0.0,0.0,26.0,0.0,14.0,32.0,14.0,0.0]],"2":[[0.76953125,0.76953125,0.76953125,0.55859375,0.46484375,0.453125,0.453125,0.39453125,0.375,0.36328125,0.3515625,0.34375,0.33203125,0.33203125,0.3203125,0.30078125,0.28125,0.28125,0.2734375,0.2734375,0.24609375,0.24609375,0.24609375,0.24609375,0.24609375,0.23828125,0.23828125,0.23046875,0.23046875,0.22265625,0.22265625,0.22265625,0.21484375,0.21484375,0.20703125,0.19921875,0.19921875,0.19921875,0.19921875,0.19140625,0.19140625,0.19140625,0.19140625,0.19140625,0.18359375,0.18359375,0.18359375,0.18359375,0.18359375,0.17578125,0.17578125,0.17578125,0.17578125,0.17578125,0.16796875,0.1640625,0.1640625,0.1640625,0.1640625,0.15625,0.15625,0.15625,0.15625,0.15625,0.15234375,0.15234375,0.15234375,0.15234375,0.14453125,0.14453125,0.14453125,0.14453125,0.14453125,0.14453125,0.140625,0.140625,0.140625,0.1328125,0.1328125,0.1328125,0.1328125,0.12890625,0.12890625,0.12890625,0.125,0.125,0.125,0.1171875,0.1171875,0.1171875,0.1171875,0.1171875,0.11328125,0.11328125,0.11328125,0.11328125,0.11328125,0.11328125,0.11328125,0.109375]],"3":[100.0]}}}}

package/python/postprocessors/yunet.py

@@ -0,0 +1,275 @@
+"""YuNet face detection postprocessor (OpenCV Zoo `face_detection_yunet`, 2023mar).
+
+Used by the `yunet-2023mar` catalog entry on the `face-detection` step through a
+per-MODEL `postprocessor: 'yunet'` override (the step default stays `scrfd`).
+
+Model contract, as published by `scripts/build-yunet-model.py`:
+
+    input   "image"   [1, 3, 640, 640]  RGB float in [0,1] — the pool's plain
+                                        `/255` tensor. Upstream YuNet wants BGR
+                                        0..255; `Mul(255)` + `Gather([2,1,0])`
+                                        are BAKED into the graph ahead of the
+                                        first Conv, identically in all three
+                                        formats, so no pool normalisation runs
+                                        (the entry declares none).
+    outputs cls_{8,16,32}  [1, N, 1]    classification probability
+            obj_{8,16,32}  [1, N, 1]    objectness probability
+            bbox_{8,16,32} [1, N, 4]    (dx, dy, log w, log h), stride units
+            kps_{8,16,32}  [1, N, 10]   5 × (dx, dy), stride units
+
+Why it cannot share `scrfd.py` (measured 2026-09-26: the SCRFD decoder finds 0
+faces on 413 person crops with YuNet's outputs):
+
+1. ONE prior per grid cell, not two; anchor-free — the prior is the cell's
+   top-left corner, `(col, row) * stride`, row-major, `N = ceil(640/s)^2`.
+2. The box is centre + size, with `exp()` on width and height — SCRFD's is four
+   distances from the anchor.
+3. The score is `sqrt(cls * obj)` over TWO tensors, both clamped to [0,1].
+
+This is a port of OpenCV's `FaceDetectorYN::postProcess`
+(`modules/objdetect/src/face_detect.cpp`).
+
+Landmarks come out in YuNet's order — right eye, left eye, nose tip, right
+mouth corner, left mouth corner, where right/left are the SUBJECT'S. That is
+image-left eye first, the same order SCRFD emits and `face-align.ts` expects
+(ARCFACE_TEMPLATE_112), so they are passed through unchanged. Measured on 196
+COCO faces: index 0 is the image-left eye except on 3 profile faces, the same
+rate as SCRFD's 2.
+
+The per-candidate arithmetic (`prior_cell`, `decode_prior`, `to_source`, `nms`)
+is plain Python so the layout, decode and landmark-order tests run on CI's bare
+interpreter; numpy only vectorises the score threshold over the 8400 priors.
+
+Output matches the scrfd postprocessor:
+``{"kind": "detections", "detections": [{"class": "face", "score", "bbox": [x1,y1,x2,y2], "landmarks"}]}``
+in ORIGINAL (pre-letterbox) pixel coordinates.
+"""
+from __future__ import annotations
+
+import math
+import sys
+import time
+
+import numpy as np
+
+STRIDES: tuple[int, ...] = (8, 16, 32)
+NUM_LANDMARKS = 5
+DEFAULT_NMS_IOU = 0.45
+
+# At most this many candidates, best first, reach the NMS. OpenCV's
+# FaceDetectorYN caps at 5000 (`top_k`), but its NMS is C++; `nms` below is
+# pure Python and O(n·kept), measured on this Mac at 28 ms for 500 disjoint
+# candidates, 113 ms for 1000 and 2.8 s for 5000 — an unbounded frame (every
+# prior passing a low floor: 8400) would hold a pool worker for ~8 s. Real
+# frames are far below the cap: on the spike's 413 person crops the most
+# candidates any frame produced was 38 at the step's 0.5 floor and 110 at 0.1.
+TOP_K = 750
+# A frame that hits the cap is reported, at most once per this interval, on
+# stderr (the pool's log): dropped work is never silent, and a pathological
+# model/threshold must not flood the log either.
+TOP_K_LOG_INTERVAL_S = 60.0
+_last_top_k_log = 0.0
+
+
+def grid_cols(stride: int, input_size: int) -> int:
+    """Priors per row (and per column — the input is square) at `stride`."""
+    return int(math.ceil(input_size / stride))
+
+
+def prior_cell(index: int, cols: int) -> tuple[int, int]:
+    """`(row, col)` of prior `index` — row-major, one prior per cell."""
+    return divmod(index, cols)
+
+
+def score_of(cls: float, obj: float) -> float:
+    """YuNet's score: geometric mean of the two probabilities, each clamped."""
+    return math.sqrt(min(max(cls, 0.0), 1.0) * min(max(obj, 0.0), 1.0))
+
+
+def decode_prior(
+    stride: int,
+    row: int,
+    col: int,
+    bbox: "list[float] | tuple[float, ...]",
+    kps: "list[float] | tuple[float, ...]",
+) -> tuple[tuple[float, float, float, float], list[tuple[float, float]]]:
+    """One prior → `(x1, y1, x2, y2)` and 5 landmark points, in NETWORK space."""
+    cx = (col + bbox[0]) * stride
+    cy = (row + bbox[1]) * stride
+    w = math.exp(bbox[2]) * stride
+    h = math.exp(bbox[3]) * stride
+    box = (cx - w / 2.0, cy - h / 2.0, cx + w / 2.0, cy + h / 2.0)
+    points = [
+        ((kps[2 * p] + col) * stride, (kps[2 * p + 1] + row) * stride)
+        for p in range(NUM_LANDMARKS)
+    ]
+    return box, points
+
+
+def to_source(x: float, y: float, scale: float, pad: tuple[int, int]) -> tuple[float, float]:
+    """Undo the pool's letterbox: network pixels → original frame pixels."""
+    return (x - pad[0]) / scale, (y - pad[1]) / scale
+
+
+def _iou(a: tuple, b: tuple) -> float:
+    ix = max(0.0, min(a[2], b[2]) - max(a[0], b[0]))
+    iy = max(0.0, min(a[3], b[3]) - max(a[1], b[1]))
+    inter = ix * iy
+    if inter == 0:
+        return 0.0
+    union = (a[2] - a[0]) * (a[3] - a[1]) + (b[2] - b[0]) * (b[3] - b[1]) - inter
+    return inter / union if union > 0 else 0.0
+
+
+def cap_top_k(candidates: list[dict], k: int = TOP_K) -> tuple[list[dict], int]:
+    """The `k` best candidates by score, and how many were dropped."""
+    if len(candidates) <= k:
+        return candidates, 0
+    best = sorted(candidates, key=lambda c: c["score"], reverse=True)
+    return best[:k], len(candidates) - k
+
+
+def _log_top_k_drop(dropped: int, total: int) -> None:
+    global _last_top_k_log
+    now = time.monotonic()
+    if now - _last_top_k_log < TOP_K_LOG_INTERVAL_S:
+        return
+    _last_top_k_log = now
+    sys.stderr.write(
+        f"yunet: {total} candidates above the confidence floor, kept the best {TOP_K} "
+        f"for NMS and dropped {dropped} (logged at most once per "
+        f"{TOP_K_LOG_INTERVAL_S:.0f}s)\n"
+    )
+    sys.stderr.flush()
+
+
+def nms(candidates: list[dict], iou_threshold: float) -> list[dict]:
+    """Greedy NMS on `_xyxy`, highest score first."""
+    kept: list[dict] = []
+    for cand in sorted(candidates, key=lambda c: c["score"], reverse=True):
+        if all(_iou(cand["_xyxy"], k["_xyxy"]) <= iou_threshold for k in kept):
+            kept.append(cand)
+    return kept
+
+
+_OUTPUT_WIDTHS: tuple[tuple[str, int], ...] = (
+    ("cls", 1), ("obj", 1), ("bbox", 4), ("kps", 2 * NUM_LANDMARKS),
+)
+
+
+def missing_outputs(output_names) -> list[str]:
+    """The named outputs this decoder needs that the model did not produce.
+
+    Matched by NAME only, with no shape fallback: `cls_*` and `obj_*` have
+    identical shapes, so a fallback could only guess. Every published format
+    names its outputs (the CoreML conversion names them explicitly, the
+    OpenVINO IR has its internal aliases cleared)."""
+    present = set(output_names)
+    return [
+        f"{kind}_{stride}" for stride in STRIDES for kind, _ in _OUTPUT_WIDTHS
+        if f"{kind}_{stride}" not in present
+    ]
+
+
+def value_count(value) -> int:
+    """Number of scalars in an ndarray-like (`.size`) or a nested sequence.
+    Pure Python, so the size gate below needs no numpy."""
+    size = getattr(value, "size", None)
+    if isinstance(size, int):
+        return size
+    if isinstance(value, (list, tuple)):
+        return sum(value_count(v) for v in value)
+    return 1
+
+
+def check_outputs(predictions: dict, input_size: int) -> None:
+    """RAISE unless all twelve outputs are present and sized for `input_size`.
+
+    Returning `[]` instead would read as "no faces" forever with no line
+    saying why — which is exactly what the SCRFD decoder did on this graph. A
+    size mismatch means the model's input is not `inputSize` (e.g. a 320
+    export under a 640 catalog entry): decoding it would put every prior in the
+    wrong cell."""
+    missing = missing_outputs(predictions.keys())
+    if missing:
+        raise ValueError(
+            f"yunet: outputs {missing} missing (model outputs: {sorted(predictions)})"
+        )
+    for stride in STRIDES:
+        priors = grid_cols(stride, input_size) ** 2
+        for kind, width in _OUTPUT_WIDTHS:
+            key = f"{kind}_{stride}"
+            got = value_count(predictions[key])
+            if got != priors * width:
+                raise ValueError(
+                    f"yunet: output '{key}' has {got} values, expected {priors * width} "
+                    f"({priors} priors × {width}) — inputSize {input_size} does not match the model"
+                )
+
+
+def _stride_tensors(predictions: dict, stride: int) -> dict:
+    """The four flattened tensors of one stride (sizes already checked)."""
+    return {
+        kind: np.asarray(predictions[f"{kind}_{stride}"], dtype=np.float32).reshape(-1)
+        for kind, _ in _OUTPUT_WIDTHS
+    }
+
+
+def postprocess_yunet(
+    predictions: dict,
+    config: dict,
+    orig_w: int,
+    orig_h: int,
+    scale: float,
+    pad: tuple[int, int],
+) -> dict:
+    """Postprocess YuNet output to structured face detections."""
+    conf_threshold = float(config.get("confidence", 0.5) or 0.5)
+    input_size = int(config.get("inputSize", 640))
+    nms_iou = float(config.get("nmsIouThreshold", DEFAULT_NMS_IOU))
+
+    check_outputs(predictions, input_size)
+
+    candidates: list[dict] = []
+    for stride in STRIDES:
+        cols = grid_cols(stride, input_size)
+        t = _stride_tensors(predictions, stride)
+        bbox = t["bbox"].reshape(-1, 4)
+        kps = t["kps"].reshape(-1, 2 * NUM_LANDMARKS)
+        # numpy only PRE-filters, conservatively: sqrt(cls·obj) <= max(cls, obj),
+        # so a prior below the threshold on both tensors cannot pass. The
+        # decision itself is `score_of`, per prior.
+        maybe = np.flatnonzero(np.maximum(t["cls"], t["obj"]) >= conf_threshold)
+        for index in maybe.tolist():
+            score = score_of(float(t["cls"][index]), float(t["obj"][index]))
+            if score < conf_threshold:
+                continue
+            row, col = prior_cell(index, cols)
+            (x1, y1, x2, y2), points = decode_prior(
+                stride, row, col, bbox[index].tolist(), kps[index].tolist()
+            )
+            sx1, sy1 = to_source(x1, y1, scale, pad)
+            sx2, sy2 = to_source(x2, y2, scale, pad)
+            ox1, oy1 = max(0.0, min(orig_w, sx1)), max(0.0, min(orig_h, sy1))
+            ox2, oy2 = max(0.0, min(orig_w, sx2)), max(0.0, min(orig_h, sy2))
+            if (ox2 - ox1) < 1 or (oy2 - oy1) < 1:
+                continue
+            landmarks = []
+            for px, py in points:
+                lx, ly = to_source(px, py, scale, pad)
+                landmarks.append({"x": round(lx, 1), "y": round(ly, 1)})
+            candidates.append({
+                "class": "face",
+                "score": round(score, 4),
+                "bbox": [round(ox1, 1), round(oy1, 1), round(ox2, 1), round(oy2, 1)],
+                "landmarks": landmarks,
+                "_xyxy": (ox1, oy1, ox2, oy2),
+            })
+
+    capped, dropped = cap_top_k(candidates)
+    if dropped:
+        _log_top_k_drop(dropped, len(candidates))
+    kept = nms(capped, nms_iou)
+    for det in kept:
+        del det["_xyxy"]
+    return {"kind": "detections", "detections": kept}