Lab geometry comparison — 20 September 2026 Input: mobile_captures/rec_20260915_150826, original 640x480 phone RGB. 128 sharp frames in equal timestamp bins, including the first and last frame. All three models use identical image names in temporal order. No geometry inputs. selection.json contains the exact original frame indices and timestamps. Workspace: /ihub/homedirs/svs_ald/sudhir/real2sim Experiment: work/mobile_lab_0915/geometry_compare Environment: envs/mapanything (Python 3.12, torch 2.7.0+cu126, NumPy 1.26.4). This venv inherits the existing sceneforge-sam3 environment; environment.txt records resolved packages. Model source revisions are in each result.json. Sources: https://github.com/facebookresearch/map-anything https://github.com/ByteDance-Seed/Depth-Anything-3 https://github.com/facebookresearch/vggt-omega Checkpoints: facebook/map-anything Hugging Face revision: a1d87e9086706fb9974f3be5a3e3a0ca5401c5aa depth-anything/DA3NESTED-GIANT-LARGE-1.1 Hugging Face revision: b2359bdf726fb44ef62acca04d629dcf158053e7 facebook/VGGT-Omega / vggt_omega_1b_512.pt Hugging Face revision: 1041e80fc0e911235d3426b0a3d9a81075111a53 The 512 model is the authors' recommended in-the-wild checkpoint. This trial does not use the 416 reproduction benchmark checkpoint. Run from the workspace root. Reuse the current holder and select an available GPU according to ~/use_instructions/README.md; the IDs below record this run. envs/mapanything/bin/python work/mobile_lab_0915/geometry_compare/prepare.py bash ~/use_instructions/run_in_holder.sh 20277 5 \ "$PWD/work/mobile_lab_0915/geometry_compare/mapanything.log" \ "$PWD/envs/mapanything/bin/python" -u \ "$PWD/work/mobile_lab_0915/geometry_compare/run.py" mapanything bash ~/use_instructions/run_in_holder.sh 20277 6 \ "$PWD/work/mobile_lab_0915/geometry_compare/omega.log" \ "$PWD/envs/mapanything/bin/python" -u \ "$PWD/work/mobile_lab_0915/geometry_compare/run.py" vggt-omega # After MapAnything completes, reuse GPU 5: bash ~/use_instructions/run_in_holder.sh 20277 5 \ "$PWD/work/mobile_lab_0915/geometry_compare/da3.log" \ "$PWD/envs/mapanything/bin/python" -u \ "$PWD/work/mobile_lab_0915/geometry_compare/run.py" da3 # After all runs complete: envs/mapanything/bin/python work/mobile_lab_0915/geometry_compare/build.py Outputs retained locally: per-view depth, confidence, RGB, camera-to-world poses, intrinsics, validity masks, and MapAnything's native world point maps. The web sample caps each view at 12,000 points, before the slider's confidence filter. Browser points retain source-view indices and confidence percentile. Both PLY downloads keep confidence >=20th percentile of the sampled cloud. The native PLY preserves the model's coordinates and scale. The aligned PLY matches the web viewer, in Y-up coordinates after a global similarity fit to the existing GLOMAP trajectory. No nonrigid adjustment or pose optimization. The reference trajectory is not ground truth. Its approximate metric scale comes from assuming 1.35 m camera height. Camera-path RMSE should only be read as disagreement with that reconstruction, not a benchmark accuracy score. The input recording is called a 103-second video in earlier reports; recorded camera timestamps span 104.704 seconds. This page uses the actual timestamps. SECOND PHONE (same lab, different recording): mobile_captures/20260920_195703.mp4 — 1920x1080, 5168 frames, 172.077 s. Experiment: work/mobile_lab_0920/geometry_compare envs/mapanything/bin/python work/mobile_lab_0915/geometry_compare/prepare_video.py \ mobile_captures/20260920_195703.mp4 --work work/mobile_lab_0920/geometry_compare Run the same run.py commands with: --work /ihub/homedirs/svs_ald/sudhir/real2sim/work/mobile_lab_0920/geometry_compare The second set used GPUs 5, 6, 7 for MapAnything, Omega, and DA3 respectively. Build with: envs/mapanything/bin/python work/mobile_lab_0915/geometry_compare/build.py \ --work work/mobile_lab_0920/geometry_compare Copy the resulting site directory to the main site's phone2/ directory. The page selects this capture with ?capture=phone2. Second-phone frames are sampled every fourth decoded frame; the sharpest in each of 128 timestamp bins is retained, with endpoints forced. Original-size images are saved as JPEG quality 95. There is no pre-existing SfM for this video. MapAnything's predicted path is used solely as the display reference, oriented using its mean camera up and a low cloud-height percentile. Other models receive one global similarity fit to that path. Its RMSE values are cross-model disagreement (MapAnything=0 by construction), not accuracy. Both input selections, exact scripts, and both sets of run metadata are published. Native full-resolution dense arrays are retained in local out/. Inference timings are CUDA-synchronized single-pass wall times. Model downloads and export are excluded. MapAnything and VGGT-Omega prepare images before the timed inference block; DA3's official inference API includes preprocessing and CPU output conversion inside that block. Peak memory is PyTorch peak allocated memory including the model. Runs used one A100-SXM4-80GB each, bf16 inference.