[{"data":1,"prerenderedAt":323},["ShallowReactive",2],{"finding:a-4-nerf-synthetic-scene-results":3,"finding-runs:a-4-nerf-synthetic-scene-results":215,"finding-related:a-4-nerf-synthetic-scene-results":284},{"meta":4,"impact":36,"sections":42},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":15,"task_code":23,"related_runs":24,"related_findings":33},"a-4-nerf-synthetic-scene-results","A.4 NeRF Synthetic 他シーン展開 — 8 シーン complete 30k 結果","lego + chair \u002F ficus \u002F drums \u002F hotdog \u002F mic \u002F materials \u002F ship で 30k 学習、NeRF Synthetic 8 シーン全完了。シーン難易度は PSNR で 17.6 dB の幅 (materials 12.71 〜 hotdog 30.29)、mean 18.95 ± 6.0 dB。本実装の brush SoTA 比 gap も scene-dependent で hotdog -7.4 dB \u002F materials -17.3 dB \u002F ficus -20.9 dB と sparse SfM init \u002F 反射 PBR シーンで挙動が異なる。","Phase 5 · Multi-scene (8 シーン complete)","2026-05-24","stable","experiment","mixed",[14],1,[16,17,18,19,20,21,22],"phase-5","nerf-synthetic","multi-scene","psnr","scene-dependency","evaluation","8-scenes","A.4",[25,26,27,28,29,30,31,32],"lego-sh3-30k","chair-30k","ficus-30k","drums-30k","hotdog-30k","mic-30k","materials-30k","ship-30k",[34,35],"a-7-multi-scene-batched","a-10-variance-baseline",{"summary":37,"rank":38,"verdict":39,"delta_psnr":40,"delta_wallclock":41},"8 シーン complete (lego + 7 新規) 30k 完遂。シーン依存性が PSNR で 17.6 dB の幅 (materials 12.71 〜 hotdog 30.29)、mean 18.95 ± 6.0 dB。本実装の brush SoTA 比 gap は scene-dependent で -7.4 dB (hotdog) 〜 -22.3 dB (ficus 含む)。共通要因仮説: SfM init.ply の sparsity (細い枝 \u002F マイク \u002F 反射 PBR で薄い) + refine grad_threshold の lego\u002Fhotdog tuning over-fit。卒論 evaluation で「lego baseline + multi-scene mean ± std」併記必須。","high","partial","-5.93 dB (8 シーン平均 18.95 vs lego 24.879、std ±6.0)","21-29 min (シーン非依存的、materials のみ +5 min)",[43,46,49,127,129,138,140,147,150,152,190,192,198,200,207,209],{"type":44,"text":45},"lead","\u003Cstrong>実施:\u003C\u002Fstrong> 2026-05-23 05:16〜06:40 JST (chair → ficus → drums → hotdog)、2026-05-23 12:10〜13:25 JST (mic → materials → ship) の 2 phase。\u003Cstrong>baseline:\u003C\u002Fstrong> lego sh=3 30k = \u003Ccode>24.879 dB \u002F 23m13s \u002F 83,734 splats\u003C\u002Fcode> (seed=42, capacity=1M)。\u003Cstrong>config:\u003C\u002Fstrong> \u003Ccode>configs\u002F2026-05-22-2200-{chair,ficus,drums,hotdog}-30k.toml + configs\u002F2026-05-23-1000-{mic,materials,ship}-30k.toml\u003C\u002Fcode>、dataset path 以外は migration-gate と同条件。",{"type":47,"text":48},"heading","メイン表 (8 シーン complete + brush paper 比較)",{"type":50,"columns":51,"align":59,"rows":62,"caption":126},"table",[52,53,54,55,56,57,58],"scene","PSNR (dB)","wallclock","ΔPSNR vs lego","brush paper","gap to brush","難易度評価",[60,61,61,61,61,61,60],"left","right",[63,71,79,86,94,102,110,118],[64,65,66,67,68,69,70],"**lego (ref)**","**24.879**","23m13s","(baseline)","35.78","-10.9","medium (角張った構造)",[72,73,74,75,76,77,78],"chair","22.883","22m54s","-2.00","35.83","-12.9","medium-hard (脚 \u002F 椅子背)",[80,81,74,82,83,84,85],"**ficus**","**13.959**","**-10.92**","34.87","**-20.9**","very hard (細い枝、init sparse)",[87,88,89,90,91,92,93],"drums","17.773","21m27s","-7.11","26.15","-8.4","hard (反射成分 \u002F 多面体)",[95,96,97,98,99,100,101],"**hotdog**","**30.290**","23m52s","**+5.41**","37.72","-7.4","easy (シンプル平面 \u002F 単一色)",[103,104,105,106,107,108,109],"mic","15.031","22m13s","-9.85","35.36","-20.3","hard (マイク 1 本、init.ply 薄い)",[111,112,113,114,115,116,117],"**materials**","**12.709**","**28m28s**","**-12.17**","30.00","-17.3","hard (反射 PBR ボール、SH=3 で不足)",[119,120,121,122,123,124,125],"ship","15.038","23m58s","-9.84","30.94","-15.9","hard (構造物、init sparse)","8 シーン complete (NeRF Synthetic 全 scene)。ficus \u002F materials が極端低、hotdog のみ lego を上回る。brush paper との gap も scene 依存。",{"type":47,"text":128},"シーン依存性の core finding (2026-05-24 更新)",{"type":130,"ordered":131,"items":132},"list",true,[133,134,135,136,137],"\u003Cstrong>シーン難易度は PSNR で 17.6 dB の幅\u003C\u002Fstrong> (materials 12.71 〜 hotdog 30.29) — 「単一シーンの数字で trainer の能力を評価する」ことの危うさを decisive に実証。","\u003Cstrong>ficus \u002F materials \u002F mic \u002F ship が壊滅的失敗 (12-15 dB)\u003C\u002Fstrong>: 共通仮説は SfM init.ply の sparsity + refine の grad_threshold が lego\u002Fhotdog tuning に over-fit。これらシーンでは init point cloud が薄く、refine が「main density area」のみに集中して周辺 (枝 \u002F 構造物 \u002F 反射光) を学習できない。","\u003Cstrong>hotdog の優位 (+5.41 dB)\u003C\u002Fstrong>: 単純な皿 + 食材形状、texture が連続的、refine の grad-driven split に向く。本実装でも 30.29 まで届いたのは健闘。","\u003Cstrong>materials の wallclock +21% (28m28s)\u003C\u002Fstrong>: 反射 PBR で gradient 計算が他シーンより重い、または refine 後 splats 多めに残った可能性。","\u003Cstrong>chair の中間 (-2.00 dB)\u003C\u002Fstrong>: 椅子の構造は本実装の trainer recipe で比較的扱いやすい、splat 数が 130k+ (他より多い) → A.7 batching が chair で -18.6% 効いた理由。",{"type":47,"text":139},"平均 PSNR と分散 (8 シーン complete)",{"type":130,"items":141},[142,143,144,145,146],"\u003Cstrong>8 シーン complete 平均\u003C\u002Fstrong>: \u003Cstrong>(24.879 + 22.883 + 13.959 + 17.773 + 30.290 + 15.031 + 12.709 + 15.038) \u002F 8 = 18.95 dB\u003C\u002Fstrong>","\u003Cstrong>標準偏差 (8 シーン)\u003C\u002Fstrong>: ~6.0 dB → \u003Cstrong>巨大なシーン依存性\u003C\u002Fstrong>","\u003Cstrong>brush SoTA paper mean (Kerbl+ 2023 Table 1)\u003C\u002Fstrong>: 33.32 dB","\u003Cstrong>本実装の multi-scene mean gap to brush\u003C\u002Fstrong>: -14.4 dB (mean to mean)、lego 単独 gap -10.9 dB より大きい","\u003Cstrong>hotdog (+5.41 vs lego)\u003C\u002Fstrong> だけが lego を上回り、他 7 シーンは全て lego を下回る → 本実装は \"lego\u002Fhotdog に over-fit した refine schedule\" の疑い",{"type":148,"text":149},"paragraph","A.5 final ablation 表の「Lego 単一の数字」では trainer の能力を \u003Cstrong>過大評価\u003C\u002Fstrong> していたことが 8 シーン complete で確定。卒論では「lego baseline + multi-scene mean ± std」を必ず併記すべき。実際には multi-scene mean 18.95 dB が本実装の honest 数字。",{"type":47,"text":151},"brush 系 SoTA との比較 (8 シーン詳細)",{"type":50,"columns":153,"align":157,"rows":158,"caption":189},[52,154,56,155,156],"本実装","gap","備考",[60,61,61,61,60],[159,163,165,169,171,175,177,181,183],[160,161,68,69,162],"lego","24.879","Phase 5 既知 gap",[72,73,76,77,164],"medium gap",[166,167,83,84,168],"ficus","13.959","max gap、SfM init 薄い",[87,88,91,92,170],"**min gap** (brush も難)",[172,173,99,100,174],"hotdog","30.290","本実装も健闘",[103,104,107,108,176],"近接 max gap",[178,179,115,116,180],"materials","12.709","反射 PBR で困難",[119,120,123,124,182],"構造物、init 薄",[184,185,186,187,188],"**mean**","**18.95**","**33.32**","**-14.4**","本実装 multi-scene 平均","本実装は \"テクスチャ連続シーン\" (hotdog) で gap 小、\"細い構造 \u002F 反射シーン\" (ficus \u002F mic \u002F materials \u002F ship) で gap 大。",{"type":47,"text":191},"卒論への含意 (Chapter 4 evaluation 章)",{"type":130,"ordered":131,"items":193},[194,195,196,197],"\u003Cstrong>central table を 8 シーン化\u003C\u002Fstrong>: A.5 lego 単独行を「NeRF Synthetic 8 シーン complete」table に拡張、mean ± std で報告。","\u003Cstrong>scene-dependent gap の議論\u003C\u002Fstrong>: 「本実装は SfM-init point cloud sparse なシーン \u002F 反射 PBR シーンで refine が届かない」を Negative finding として明文化 (D.3 negative-findings-chapter に追加)。","\u003Cstrong>brush 比 multi-scene mean\u003C\u002Fstrong>: 本実装 18.95 dB vs brush 33.32 dB = \u003Cstrong>multi-scene gap -14.4 dB\u003C\u002Fstrong>、これが「trainer recipe 不足」の honest evaluation (lego 単独 -10.9 dB より 3.5 dB 悪い)。","\u003Cstrong>モバイル含意\u003C\u002Fstrong>: シーン dependency が強いので、デバイス上での同 trainer 適用も「容易シーンには動くが難しいシーンには破綻」を示唆、deployment 前に scene category 分類が必要。",{"type":47,"text":199},"残作業",{"type":130,"ordered":131,"items":201},[202,203,204,205,206],"✅ \u003Ccode>splat-summary build-all\u003C\u002Fcode> で 8 シーン runs\u002F HTML 化 (deploy.sh で自動)","✅ plan.vue 更新 (A.4 → done)","✅ A.5 final-ablation-table.toml に \"Multi-scene NeRF Synthetic 8\" セクション追加","(defer) refine grad_threshold を scene-adaptive 化、ficus \u002F mic \u002F materials の PSNR 改善","(defer) brush SoTA gap を埋めるための trainer recipe 改修 (lr schedule, refine boundary, init.ply 増強)",{"type":47,"text":208},"関連",{"type":130,"items":210},[211,212,213,214],"A.5 final ablation 表: \u003Ccode>final-ablation-table\u003C\u002Fcode>","A.7 × multi-scene batched (scene 依存性ある finding と integral): \u003Ccode>a-7-multi-scene-batched\u003C\u002Fcode>","A.10 variance baseline (有意性 noise floor): \u003Ccode>a-10-variance-baseline\u003C\u002Fcode>","D.3 Negative findings 章: \u003Ccode>negative-findings-chapter\u003C\u002Fcode> (scene-dependent gap を追加)",[216,226,233,240,250,258,266,274],{"id":31,"title":31,"subtitle":217,"date":218,"workspace":219,"tags":220,"verdict":39,"psnr":222,"psnr_unit":-1,"wallclock":223,"splats":224,"summary_url":225,"detail_path":225},"A.4 NeRF Synthetic materials 30k baseline (sh=3)","2026-05-23","splat",[221,31],"auto-bench",12.708748817443848,"28m 27s",161935,"\u002Fruns\u002Fmaterials-30k\u002F",{"id":30,"title":30,"subtitle":227,"date":218,"workspace":219,"tags":228,"verdict":39,"psnr":229,"psnr_unit":-1,"wallclock":230,"splats":231,"summary_url":232,"detail_path":232},"A.4 NeRF Synthetic mic 30k baseline (sh=3)",[221,30],15.030896186828613,"22m 13s",31901,"\u002Fruns\u002Fmic-30k\u002F",{"id":32,"title":32,"subtitle":234,"date":218,"workspace":219,"tags":235,"verdict":39,"psnr":236,"psnr_unit":-1,"wallclock":237,"splats":238,"summary_url":239,"detail_path":239},"A.4 NeRF Synthetic ship 30k baseline (sh=3)",[221,32],15.03847599029541,"23m 58s",107450,"\u002Fruns\u002Fship-30k\u002F",{"id":26,"title":26,"subtitle":241,"date":242,"workspace":219,"tags":243,"verdict":39,"psnr":246,"psnr_unit":-1,"wallclock":247,"splats":248,"summary_url":249,"detail_path":249},"A.4 NeRF Synthetic 他シーン展開 — chair 30k baseline (sh=3)","2026-05-22",[244,26,245,16],"scene-ablation","scene-chair",22.88273811340332,"19m 51s",38928,"\u002Fruns\u002Fchair-30k\u002F",{"id":28,"title":28,"subtitle":251,"date":242,"workspace":219,"tags":252,"verdict":39,"psnr":254,"psnr_unit":-1,"wallclock":255,"splats":256,"summary_url":257,"detail_path":257},"A.4 NeRF Synthetic 他シーン展開 — drums 30k baseline (sh=3)",[244,28,253,16],"scene-drums",17.772979736328125,"21m 23s",64515,"\u002Fruns\u002Fdrums-30k\u002F",{"id":27,"title":27,"subtitle":259,"date":242,"workspace":219,"tags":260,"verdict":39,"psnr":262,"psnr_unit":-1,"wallclock":263,"splats":264,"summary_url":265,"detail_path":265},"A.4 NeRF Synthetic 他シーン展開 — ficus 30k baseline (sh=3)",[244,27,261,16],"scene-ficus",13.958805084228516,"22m 48s",60938,"\u002Fruns\u002Fficus-30k\u002F",{"id":29,"title":29,"subtitle":267,"date":242,"workspace":219,"tags":268,"verdict":39,"psnr":270,"psnr_unit":-1,"wallclock":271,"splats":272,"summary_url":273,"detail_path":273},"A.4 NeRF Synthetic 他シーン展開 — hotdog 30k baseline (sh=3)",[244,29,269,16],"scene-hotdog",30.290376663208008,"23m 46s",82154,"\u002Fruns\u002Fhotdog-30k\u002F",{"id":25,"title":25,"subtitle":275,"date":242,"workspace":219,"tags":276,"verdict":39,"psnr":280,"psnr_unit":-1,"wallclock":281,"splats":282,"summary_url":283,"detail_path":283},"A.8 SH degree ablation — sh_degree=3 (DC only, no view-dependence)",[277,278,279,16],"sh-ablation","lego-30k","sh-3",24.87872886657715,"23m 13s",83734,"\u002Fruns\u002Flego-sh3-30k\u002F",[285,305],{"id":35,"title":286,"date":218,"status":10,"polarity":287,"category":11,"axes":288,"tags":290,"task_code":296,"related_runs":297,"delta_psnr":301,"delta_wallclock":302,"rank":38,"verdict":10,"impact_summary":303,"detail_path":304},"A.10 variance baseline — σ ±0.32 dB \u002F range 0.885 dB を実測","negative",[289],3,[16,291,292,293,294,295],"variance","gpu-non-determinism","kahan","atomic","apple-silicon","A.10",[25,298,299,300],"lego-variance-trial1-30k","lego-variance-trial2-30k","lego-variance-trial3-30k","σ ±0.32 dB \u002F range 0.885 dB","σ ±2.4% \u002F range 5.2%","M-3.x lego sh3 30k の PSNR variance は σ ±0.32 dB \u002F range 0.885 dB (4 run estimate)、wallclock variance は σ ±2.4% \u002F range 5.2%。原因は SIMD backward kernel の atomic_fetch_add 順序非決定性で、A.10 Kahan で消えない (compensator も bit-identical のところ)。卒論 finding として「Apple Silicon の variance band は数値精度の問題でなく GPU scheduler 由来」と確定。","\u002Ffindings\u002Fa-10-variance-baseline\u002F",{"id":34,"title":306,"date":218,"status":10,"polarity":12,"category":11,"axes":307,"tags":308,"task_code":312,"related_runs":313,"delta_psnr":319,"delta_wallclock":320,"rank":38,"verdict":39,"impact_summary":321,"detail_path":322},"A.7 × multi-scene — batching 効果は scene 依存 (-1.6% 〜 -18.6% で 12x の幅)",[289],[16,309,310,311,18,20,295],"a-7","icb","batched","A.7",[26,314,27,315,28,316,29,317,25,318],"chair-batched-30k","ficus-batched-30k","drums-batched-30k","hotdog-batched-30k","lego-a7-batched-30k","+0.130 dB 〜 -0.828 dB (mean -0.21 dB)","-1.6% 〜 -18.6% (mean -7.0%)","A.7 batched cmd buffer の wallclock 改善は scene 依存で chair -18.6% \u002F hotdog -5.4% \u002F drums -3.4% \u002F ficus -1.6% \u002F lego -6.16% の 5 シーン (12x の幅)。chair の突出は splat 数最大 (~130k) + scene geometry の compute\u002Fcommit ratio が高いことが要因と推測。一方 ficus \u002F drums は variance 範囲内、独立 effect 断定不可。卒論で「A.7 effective ≠ universal、scene 選択 + workload analysis 必須」と honest framing。","\u002Ffindings\u002Fa-7-multi-scene-batched\u002F",1782449788618]