code/docs/BASELINES.md

# UniCal baselines for KITTI-C2L v1

These historical training results use the frozen v1.0 x-only calibration convention. See [the v1.1 calibration guide](CALIBRATION.md) for the additive full-projection protocol and re-evaluation of these same checkpoints.

Download the [verified checkpoints, source and evaluation harness](https://huggingface.co/datasets/mcocheteux/KITTI-C2L-Dataset/resolve/v1.0.0/baselines/kitti-c2l-unical-baselines-v1.0.0.tar.gz) (73 MB). The archive includes its licenses and checksums.

These results establish useful learning on the frozen data. They do not
reproduce the published UniCal training recipe or advertised accuracy. All
values below average seeds 42, 43 and 44; translation is MAE in centimetres and
rotation is mean geodesic degrees. Lower is better.

## Longer unchanged-main baseline

Model implementation: UniCal main commit
`7c16fd81e4889d6f16076b1ec5b58835d876807b`. Each C2L run completed 50 epochs,
925,350 encoder frames and 115,700 updates. Checkpoints were selected by
sequence-07 validation loss at epochs 45/34/40 (one-based).

| Held-out target protocol | Identity correction | C2L-trained UniCal | Usual-KITTI-trained UniCal |
| --- | ---: | ---: | ---: |
| Native frozen C2L | 6.6925 / 1.3287 | 4.0628 / 0.7746 | 7.2228 / 1.3284 |
| Usual-relative control | 5.0536 / 0.9542 | 3.5458 / 0.6061 | 5.6084 / 0.9543 |

C2L improves both metrics versus identity and the usual-KITTI comparator,
overall and separately on each held-out drive. The usual comparator remains
weak, so this is evidence that C2L is learnable, not evidence of matching a
strong published baseline. Its runs used 61 epochs, 933,666 encoder frames and
116,754 updates (approximately 0.9% more exposure). Validation curves were
non-monotonic; these are validation-selected checkpoints.

Recipe: 256×256 inputs, ImageNet RGB normalization, 8,192 spatial-loss points
sampled **after dense projection**, BF16 mixed precision, batch 8, four spawned
loader workers, pinned memory, uniform learning rate 3e-5 and one warmup epoch.
The model uses pretrained MobileViT-small and 6D rotation outputs. This recipe
is different from the paper's grayscale/RPY recipe and training duration.

## Perturbation-policy comparison

Under the matched short recipe on the improved implementation, native errors
were 4.2528 / 0.6148 for frozen C2L versus 4.4919 / 0.6524 for usual KITTI.
On usual-relative targets they were 3.7726 / 0.5199 versus 3.2897 / 0.4555:
**14.68% translation and 14.14% rotation deficits**, both beyond the study's
10% margin, with the same finding on each test drive.

Regenerating usual-style online perturbations on C2L frames gave
3.2632 / 0.4368 on usual-relative targets and met the 10% comparison criterion
on both protocols. These results support the frame/GT content and show that
the frozen perturbation distribution affects transfer. They do not establish
universal parity for frozen C2L.

Three-frame inference averaging improved frozen-C2L results from
4.3470 / 0.6343 to 4.1036 / 0.5944 on 2,604 common window anchors. Dedicated
three-frame training did not establish an accuracy gain and used three times
the encoder exposure. Keep those experiments distinct.

Full design, uncertainty and limitations: [ASSESSMENT.md](ASSESSMENT.md).
Compact numerical evidence and original report hashes:
[assessment_results.json](../release/v1/assessment_results.json).

## Reproducing the packaged baseline

The preparation script can create a **separate** local baseline package with
three C2L-trained checkpoints. Exports contain plain CPU tensors, epoch and
update counts, with no optimizer state or serialized Python objects. The
manifest records the original checkpoint hashes and exported hashes.

The package includes a frozen UniCal source archive at main commit `eebb3b0`
(full hash in its manifest), with the same model/preprocessor as the training
baseline and the independently usable loader fix. It also includes the
experiment harness source at `2550c6d`, its study protocol, and this release's
offline-capable native evaluator. The verified archive is publicly downloadable through the link above.

After unpacking the baseline package, extract both `unical-source.tar` and
`harness-source.tar`. Install the UniCal
source with `pip install /path/to/unical`. For the pinned dependencies, use
`uv sync --project /path/to/unical --python 3.11 --locked` and run with that
environment's Python. Run the included evaluator:

```bash
python evaluate_baseline.py \
  --checkpoint main-c2l-long50-fast256-seed42/checkpoints/model.pt \
  --data-dir /path/to/kitti-c2l/data \
  --raw-root /path/to/kitti_raw \
  --device cpu --output seed42-native.json
```

Use `--device cuda` when available. Evaluate all three seeds for the reported
mean. `--max-samples 2` is a smoke check and is not a benchmark result.
If the source is used without installation, set `PYTHONPATH=/path/to/unical/src`.
Model construction disables pretrained downloads because the export contains
all weights. The native evaluator uses stage 5, 256px and ImageNet RGB; it does
not regenerate targets. The original experiment evaluator in the harness
supports the secondary relative protocol and paired analyses. Its GPU path
requires local CUDA; pass paths explicitly instead of its historical defaults.

The original evaluator used batches of two; the packaged evaluator keeps that
default to minimize GPU-kernel rounding differences. Different devices, batch
sizes and library versions can produce small numeric differences.

The training harness records Git provenance. Use Git checkouts at the recorded
revisions, or initialize and commit each extracted `unical` and `harness` source
directory locally before new training. Snapshot commits will have new IDs;
the source-archive hashes preserve their original provenance.

For new training, install the source archive and use the included harness
with that source on PYTHONPATH:

```bash
PYTHONPATH=/path/to/unical/src python harness/experiments/run_comparison.py \
  --variant main --protocol frames --stage 5 --seed 42 \
  --data-dir /path/to/kitti-c2l/data --raw-root /path/to/kitti_raw \
  --epochs 50 --warmup-epochs 1 --batch-size 8 --precision bf16-mixed \
  --image-size 256 --max-loss-points 8192 --num-workers 4 --pin-memory \
  --validation-period 1 \
  --output /path/to/new-run
```

Use seeds 43/44 for repeats. The harness installs no dataset and performs no
upload. A basic `train.py data=kitti_c2l` command uses different defaults and
must not be presented as exact reproduction of these numbers.

## Licensing

Dataset metadata is CC BY-NC-SA 3.0. UniCal code and the separately packaged
model artifacts are supplied under CC BY-NC 4.0, with attribution and source
license included; these files are not relicensed as MIT or as dataset metadata.
The backbone originates from `apple/mobilevit-small`, whose model card links
to the Apple sample code license. That upstream notice is retained separately
and is not replaced by CC BY-NC 4.0; see THIRD_PARTY.md. No training images or point
clouds are bundled with either package. Public release of the baseline package
is a separate publication step.