5^th Workshop on Maritime Computer Vision (MaCVi)

MaCVi @ WACV 2027

Tuesday, January 5, 2027, afternoon (Orlando local time), per WACV’s draft schedule. Exact times and room to be confirmed.

Challenges / Maritime Visual Odometry

Maritime Visual Odometry

Estimate a boat-mounted camera's trajectory from RGB or thermal video.

Challenge and dataset supported by LOOKOUT.

Challenge schedule · Teams & paper eligibility · Release notes

Task and inputs

Predict a timestamped camera-position trajectory for each monocular clip, using only video, camera calibration and decoded frame timestamps. GPS, IMU, AIS, speed, heading and other clip-specific navigation signals are excluded. Reference trajectories are evaluation labels.

Data and setup

The development set has 68 training and 24 validation clips, split by recording boat, with RGB and thermal video. It includes frame timestamps, nominal field-of-view-derived calibration and separate reference positions in metres (local origin, zero vertical coordinate).

Development data (6.8 GB) SHA-256 checksum

The test set has 20 clips with video, calibration and timestamps. Test reference trajectories remain private.

Test data (1.3 GB) SHA-256 checksum

Use the public starter for download verification and a synthetic evaluator example (Python and NumPy, no GPU). The setup guide covers the DPVO baseline, CUDA environment, inference and local evaluation. Use the repository's current starter rather than the older snapshot in the archives.

Predictions and scoring

Each clip's JSON prediction contains key, frame_indices, times and finite 3D positions in one coordinate system. Indices and timestamps must match the released video frames. See the prediction format.

Scoring aligns each trajectory with one global similarity transform, including scale, and measures mean relative-position drift as a percentage of segment length over eligible 100–800 metre segments. Lower drift is better; orientation and absolute scale are not scored. Absolute trajectory error is also reported. Report scored, failed and missing clips and temporal coverage alongside drift; drift statistics cover scored clips only. Development scores are not competition entries.

Test submissions

Submit one JSON prediction bundle, not a model, for all 20 test clips (maximum 64 MB). Run inference on your own hardware, then follow the packing instructions to create submission.json.

1. Create or join a team on Teams & paper. The team owner uploads predictions. Author details may be completed by the separate author-metadata deadline; completing them does not guarantee paper inclusion.
2. Upload the bundle and inspect the evaluation report. All 20 clips must score, each with at least 95% temporal coverage. Coverage counts reference timestamps supported by a predicted sample or interpolation across gaps of at most 1 second. Long gaps and endpoint-only trajectories do not qualify.
3. Select one eligible upload as your final entry in the dashboard to appear on the leaderboard. You can change it until the submission deadline.

One evaluated submission per team and uploader per UTC day (reset at 00:00 UTC). Pending evaluations reserve the slot. Invalid-format failures release it; scored but incomplete or ineligible entries consume it. Deletion does not restore the allowance. Team attribution is fixed at upload time.

Training and terms

Train on the supplied training split. Disclose external pretraining. Validation references may be used for development and evaluation; do not train on validation or test videos, including self-supervised training. The dataset uses CC BY-NC 4.0; starter code uses MIT. Upstream DPVO and checkpoint terms apply separately. Read the supplied LICENSE.txt and full challenge rules.

Updates and support

Ask questions in the visual-odometry-support-forum (join MaCVi Discord). Check release notes and known limitations for starter updates.