[{"data":1,"prerenderedAt":135},["ShallowReactive",2],{"finding:a-11-tanks-temples-investigation":3,"finding-runs:a-11-tanks-temples-investigation":105,"finding-related:a-11-tanks-temples-investigation":106},{"meta":4,"impact":27,"sections":31},{"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":25},"a-11-tanks-temples-investigation","A.11 Tanks & Temples real-world シーン — 実装試行せず documented investigation で close","Tanks & Temples (Truck \u002F Barn) は dataset DL 5-10GB + COLMAP 前処理 + 複数 long run が必要で卒論 time-box に合わず。NeRF Synthetic 8 シーン (A.4 で 7 新規追加) で evaluation の幅出しは十分。","Investigation · Real-world dataset","2026-05-23","stable","spec","negative",[14],1,[16,17,18,19,20,21,22],"phase-5","tanks-and-temples","real-world","colmap","investigation","deferred","nerf-synthetic","A.11",[],[26],"a-4-nerf-synthetic-scene-results",{"summary":28,"rank":29,"verdict":30},"T&T を取り込むには dataset DL 5-10GB + COLMAP SfM + trainer 入力 format 対応 + eval indicator 再設計 + long run × 数シーンが必要で 2-3 日コース。NeRF Synthetic 8 シーン整備で幅出しは達成済みのため close。","mid","investigative",[32,35,40,43,46,52,54,60,62,70,72,74,76,81,83,85,92,94,96,99,101],{"type":33,"text":34},"lead","Tanks & Temples (T&T) の real-world シーン (Truck \u002F Barn) は \u003Cstrong>dataset DL 5-10GB + COLMAP 前処理 + 複数 long run\u003C\u002Fstrong> が必要で、卒論完成までの時間枠で見合わない。NeRF Synthetic 8 シーン (A.4 で 7 新規追加) で evaluation の幅出しは十分達成。",{"type":36,"label":37,"variant":38,"text":39},"callout","ユーザー判断 (2026-05-23)","info","A.11 は investigation doc で close、実装はしない。",{"type":41,"text":42},"heading","なぜ scope 大か",{"type":44,"text":45},"paragraph","NeRF Synthetic は合成データなので:",{"type":47,"items":48},"list",[49,50,51],"100 train + 100 val + 200 test views、800x800 px、background pure white","camera intrinsics が \u003Ccode>transforms_train.json\u003C\u002Fcode> で完備、focal length が exact","\u003Ccode>init.ply\u003C\u002Fcode> (sparse point cloud, ~10000 vertex) が既存 in dataset",{"type":44,"text":53},"T&T は real-world dataset で:",{"type":47,"items":55},[56,57,58,59],"shoot された写真 (RGB + EXIF metadata)、解像度・F 値・焦点距離が動的","camera pose は COLMAP の SfM (Structure-from-Motion) で別途取得必要","background が real (sky \u002F wall \u002F clutter)、α-mask が無いので white-bg eval が成立しない","init point cloud も COLMAP の sparse cloud (合計画像数 × keypoint 数 ≫ 10000)",{"type":44,"text":61},"つまり、T&T で run するには:",{"type":47,"ordered":63,"items":64},true,[65,66,67,68,69],"\u003Cstrong>dataset DL\u003C\u002Fstrong>: ~5-10 GB (Truck\u002FBarn だけでも 2 シーン × 数 GB)","\u003Cstrong>COLMAP 前処理\u003C\u002Fstrong>: SfM 実行 (CPU ベース、数十分〜数時間)、camera pose と sparse cloud 取得","\u003Cstrong>trainer 入力 format 対応\u003C\u002Fstrong>: 現状 \u003Ccode>splat-io\u002Fdataset.rs\u003C\u002Fcode> は \u003Ccode>load_nerf_synthetic\u003C\u002Fcode> のみ、T&T 用 \u003Ccode>load_tanks_and_temples\u003C\u002Fcode> を新規実装する必要","\u003Cstrong>eval convention\u003C\u002Fstrong>: white-bg eval ができないので val PSNR 比較も新指標 (e.g. SSIM、LPIPS、PSNR over masked region)","\u003Cstrong>long run\u003C\u002Fstrong>: 1 シーン 30k で 30-50 min (capacity 大)、4 シーンで 2-3 時間",{"type":44,"text":71},"これは 2-3 日かかる作業で、卒論 central table の 1 row 追加に見合わない。",{"type":41,"text":73},"NeRF Synthetic 8 シーンで evaluation の幅出しは十分",{"type":44,"text":75},"A.4 で chair \u002F ficus \u002F drums \u002F hotdog \u002F mic \u002F materials \u002F ship + 既存 lego = 8 シーン全揃った。これで:",{"type":47,"items":77},[78,79,80],"幾何複雑度の幅 (chair: 複雑、ficus: 細い枝、drums: 反射、hotdog: 凹凸、ship: 構造物)","シーンサイズの幅 (mic: 小、ship: 大)","material 複雑度の幅 (drums: brushed metal, materials: PBR ball)",{"type":44,"text":82},"→ シーン依存性の議論は十分。real-world は卒研範囲外と切る。",{"type":41,"text":84},"実装するなら必要な手順",{"type":47,"ordered":63,"items":86},[87,88,89,90,91],"T&T 公式 DL link (\u003Ccode>https:\u002F\u002Fwww.tanksandtemples.org\u002Fdownload\u002F\u003C\u002Fcode> の \"Training Data\"、Truck\u002FBarn など) で dataset 取得","COLMAP (Apple Silicon 対応 brew package あり) で SfM 実行","\u003Ccode>splat-io\u003C\u002Fcode> に \u003Ccode>load_tanks_and_temples\u003C\u002Fcode> を追加","eval indicator を masked-PSNR \u002F SSIM \u002F LPIPS で再設計","30k bench × 2-4 シーン",{"type":41,"text":93},"卒論への含意",{"type":44,"text":95},"A.5 final ablation 表 + central comparison は NeRF Synthetic に絞る。卒論本文に「real-world data (T&T) は dataset 整備の前処理コスト + eval convention の再設計が必要で、本研究の time-box 内では NeRF Synthetic に限定した。今後の研究課題」と書く。",{"type":36,"label":97,"variant":38,"text":98},"Stretch goal","卒研後の追研究 \u002F 修論で再着手候補。",{"type":41,"text":100},"関連",{"type":47,"items":102},[103,104],"A.4 NeRF Synthetic 8 シーン展開: chair-30k 進行中 (2026-05-23)、4 シーン完遂で central table の幅出し","A.5 final ablation 表 第 1 軸: \u003Ccode>final-ablation-table.md\u003C\u002Fcode>",[],[107],{"id":26,"title":108,"date":109,"status":10,"polarity":110,"category":111,"axes":112,"tags":113,"task_code":119,"related_runs":120,"delta_psnr":129,"delta_wallclock":130,"rank":131,"verdict":132,"impact_summary":133,"detail_path":134},"A.4 NeRF Synthetic 他シーン展開 — 8 シーン complete 30k 結果","2026-05-24","mixed","experiment",[14],[16,22,114,115,116,117,118],"multi-scene","psnr","scene-dependency","evaluation","8-scenes","A.4",[121,122,123,124,125,126,127,128],"lego-sh3-30k","chair-30k","ficus-30k","drums-30k","hotdog-30k","mic-30k","materials-30k","ship-30k","-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)","high","partial","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。","\u002Ffindings\u002Fa-4-nerf-synthetic-scene-results\u002F",1782449788618]