UW-IPL TideGraph - entry 19406 artifacts
Public organizer-review artifacts for the reported 10.83% entry in the MaCVi 2027 Maritime Visual Odometry challenge. Code: https://github.com/heejaeee/TideGraph-MaCVi2027-VO
Downloadable checkpoints
checkpoints/TideGraph-MaCVi2027-VO_checkpoints_verified.zip is a rebuilt, CRC-tested bundle containing:
- VGGT-Omega 1B 416-reproduce (also used at 512 input resolution).
- Pi3X pretrained weights.
- DPVO pretrained weights.
- UW-IPL flow_speed_v3 regressor and its TRAIN68 training record.
Every checkpoint is unmodified and checked by SHA-256 against the original project's pinned weights. The supplied companion ZIP could not be opened; this valid replacement was built from the original local weight files, with licenses and attribution added. checkpoints/verification.json records size, ZIP CRC validation and hashes. The original invalid ZIP is not published.
This is the four-model companion bundle, not all raw-pipeline assets. Additional R3, DA3 metric, VGGT-1B, AnyCalib, CoTracker3 and Mask2Former weights are obtained by scripts/setup_raw_assets.py in the code repository. That script accepts --checkpoint-zip to use this bundle for the four weights above. TRAIN/VAL videos are not needed to run the trained inference pipeline.
Other artifacts
models/flow_speed_v3.pkl: the team's HistGradientBoosting speed regressor fitted only on the official 68 TRAIN clips, scikit-learn 1.7.2.artifacts/estimated_signals_entry19406.zip: inferred Omega-window rotations and retained speed/heading signals for exact cached replay. These are estimates, not ground truth.ARTIFACT_MANIFEST.json: hashes for the original cached-release assets.
Verification scope
Replay/rebuild of archived estimates reproduces the submitted entry JSON exactly. The isolated raw_run.py path has now completed all 20 trajectories from released TEST videos and frozen weights, without reading archived predictions or navigation references. Timestamps and frame indices match, with 100% temporal coverage. Fresh GPU output is not byte-identical: prediction-to-prediction segment-drift self-consistency is 0.005156% mean and 0.081205% maximum. These values compare estimates in their arbitrary trajectory units; they are not a new hidden-TEST score. The official reported entry remains #19406 at 10.83%.
Full evidence: https://github.com/heejaeee/TideGraph-MaCVi2027-VO/blob/main/evidence/raw_video_all20.json Raw setup and inference: https://github.com/heejaeee/TideGraph-MaCVi2027-VO/blob/main/RAW_REPRODUCTION.md Method: https://github.com/heejaeee/TideGraph-MaCVi2027-VO/blob/main/METHOD.md The factsheet, author contact details and organizer correspondence are supplied privately by the team.
A second full raw run in the separate Python environment and a fresh public clone completed all 20 clips (180 stages), with 100% coverage. Prediction-to-prediction drift was 0.004841% mean and 0.081418% maximum. This used the same host/GPU, not a clean operating system, and reused verified weights but no estimated intermediates from the first run. The missing MoviePy dependency exposed by this run is fixed in the installation requirements. Frozen code and installation revision: 87d888d2b73da63fa98284eddd877a14ba1980f7. See evidence/independent_raw_all20.json in the code repository. These are reproduction comparisons, not official rescoring.
Component-specific terms
Noncommercial research use only for the checkpoint bundle. There is no blanket MIT/Apache license for all models.
- VGGT-Omega: FAIR Noncommercial Research License and its Acceptable Use Policy. Its Section 1 permits redistribution under that agreement when its full copy is provided. The agreement is included in the ZIP and under
checkpoints/licenses/. Original source: https://huggingface.co/facebook/VGGT-Omega . The upstream download uses manual gating; this mirror does not change its license or acceptable-use obligations. - Pi3X weights: CC BY-NC 4.0, as stated in https://huggingface.co/yyfz233/Pi3X . Pi3 source code is separately BSD licensed. Both notices are included.
- DPVO: MIT, Copyright 2022 Princeton Vision & Learning Lab; notice included. https://github.com/princeton-vl/DPVO .
- UW-IPL speed regressor: team-trained on the released TRAIN split; applicable dataset terms remain in effect.
Retain original authors' attribution and component terms when sharing. No original video, reference trajectory, access token or credential is included. Dataset metadata/derived artifacts retain applicable LOOKOUT CC BY-NC 4.0 terms. This repository does not relicense third-party material.
Team: UW-IPL, University of Washington. Organizer contact details are supplied privately.