pash03023/TartanIMU-CoCEL_POSTECH

TartanIMU Challenge — CoCEL@POSTECH Final Submission

Single unified model for IMU-only body-frame velocity estimation across four platforms (car / quadruped / drone / human). Leaderboard public 0.25392.

Contents

File	Description
CoCEL_POSTECH.pt	Frozen weights of the single unified model (7,979,267 params, one weight set)
predict.py	Self-contained inference entry point (no internet, no external repo imports)
submission_CoCEL_POSTECH.csv	The exact prediction file that produced the claimed leaderboard score
requirements.txt	Pinned dependencies (torch / numpy / pandas / scipy only)
src/genz_imu/	Model, encoders, batching, windowing, and the training script
ref/TartanIMU/	The organizers' metric module, unmodified — local scoring only

No normalisation constants file: the model is trained on the raw signal, and every preprocessing setting it needs is stored inside the checkpoint.

How to Inference

pip install -r requirements.txt
python predict.py --data /path/to/challenge_data \
                  --weights CoCEL_POSTECH.pt --out submission_CoCEL_POSTECH.csv

--data is the split root: the directory holding test/, index/ and sample_submission.csv. Output is 30,644 rows in 18 s on one RTX 5090.

⚠ CUDA is required to reproduce submission_CoCEL_POSTECH.csv exactly. requirements.txt pins the CUDA build of torch that produced it, so the pip install above fetches it; a CPU-only wheel computes the same function with different floating-point sums and does not match.

Notes

• One model, one weight set. Every prediction for all four platforms comes from the single state_dict in CoCEL_POSTECH.pt. No ensemble, no weight averaging, no per-platform checkpoint.

Contact

• Sanghyun Park: pash0302@postech.ac.kr
• Seongjun Kim: seongjun4527@postech.ac.kr