Datasets:

mcocheteux
/

KITTI-C2L-Dataset

# Calibration provenance and the v1.1 projection correction

KITTI provides a carefully calibrated reference, with measurement uncertainty.
Our v1.1 update corrects how KITTI-C2L converts that reference into a rectified
camera-2 transform. It does not claim a new, independently measured calibration.

## What KITTI originally calibrated

The [KITTI data paper](https://www.cvlibs.net/publications/Geiger2013IJRR.pdf)
describes daily calibration after recording. Camera–LiDAR registration used
50 manually selected correspondences and a robust stereo-disparity criterion.
It also documents the spinning LiDAR's acquisition times and camera triggering.
Those procedures support using KITTI as a reference, but do not establish a
sub-centimetre uncertainty bound for every driving frame.

[Subsequent camera recalibration work](https://arxiv.org/abs/2109.03462)
reports improvements in odometry. That is evidence that calibration can be
refined for a particular task, rather than proof of a better camera–LiDAR
extrinsic ground truth for KITTI-C2L.

## Verified correction

For a LiDAR point in metres, KITTI's documented projection is

```text
image_h = P_rect_02 @ R_rect_00 @ T_cam0_from_velodyne @ point_h
```

Writing `P = [K | p]` gives `P = K [I | b]`, where `b = solve(K, p)`.
Consequently, the transform consumed by a loader that applies `K` separately is

```text
R = R_rect_00 @ R_cam0_from_velodyne
t = R_rect_00 @ t_cam0_from_velodyne + solve(K, P[:,3])
```

The numeric source files have non-zero second and third entries in `P[:,3]`.
Keeping only `P[0,3]/P[0,0]`, as frozen v1.0 does, discards those entries and the
principal-point coupling in the x translation. The difference is:

| Capture date | Full minus legacy translation, cm (x, y, z) | Euclidean difference |
| --- | --- | ---: |
| 2011-09-30 | −0.528028, −0.143972, +0.620322 | 0.827249 cm |
| 2011-10-03 | −0.319263, −0.113119, +0.377976 | 0.507534 cm |

The earlier v1 documentation's “approximately 0.62 cm” described the largest
coordinate difference, not the full vector norm. Use the norms above when
describing the total translation difference.

This is an algebraic consistency correction. It does not estimate sensor poses
from images, fit test sequences, alter raw calibration files, or show that the
physical KITTI calibration itself was wrong by these amounts.

## Evidence and scope

The [machine-readable audit](../release/v1.1.0/calibration_audit.json) records
source calibration hashes, matrices, changes and every regenerated file hash.
The audit uses three predetermined frames per drive (10%, 50%, 90% of the
sequence), over all ten included drives. Points are restricted to the original
image field of view and projected depths above three metres.

- 30 frames and 590,813 projected LiDAR points were checked.
- Legacy conversion differs from the supplied projection by 0.22–0.50 pixels
  on average per drive, with per-drive 95th percentiles of 0.54–1.29 pixels.
- The full conversion in float64 has a maximum discrepancy below 1e-12 pixels.
- Actual stored float32 calibration fields have a maximum discrepancy below
  0.000039 pixels on the audited points.
- Regeneration reproduced all four corrected Parquet files byte-for-byte.
- All rows retain the same IDs, seeds, deltas, rotations, intrinsics, source
  mappings and splits. Only GT and miscalibrated translations change; the
  latter are recomposed, not edited independently.

These are projection-consistency checks against the supplied KITTI matrix,
not errors against manually observed image correspondences or independent
physical measurements. The source projection is the reference in this test.

The six original calibration text files are included unchanged under
`release/v1.1.0/source_calibration/`, so source consistency can be checked
without downloading any sensor recordings.

The [standalone reference matrices](../release/v1.1.0/calibration_reference.json)
provide both metre and centimetre translations, the camera intrinsics and the
original source-file hashes.

## Choosing a protocol

| Hugging Face configuration | Calibration convention |
| --- | --- |
| `metadata`, `windows` | Frozen v1.0, x-only camera shift; unchanged default |
| `metadata_full_projection`, `windows_full_projection` | v1.1, complete numeric projection; recommended for new experiments |
| `frames` | Shared immutable KITTI sensor bytes; independent of label convention |

```python
from datasets import load_dataset
repo = "mcocheteux/KITTI-C2L-Dataset"
metadata = load_dataset(repo, "metadata_full_projection", revision="v1.1.0")
windows = load_dataset(repo, "windows_full_projection", revision="v1.1.0")
```

The existing loader materializes both conventions from their respective rows.
Never mix GT from one convention with perturbed transforms from the other.
Use one convention consistently for training, validation and testing, and name
it in results. The original `v1.0.0` tag and sensor shards remain unchanged.

The [new checkpoint evaluation](../release/v1.1.0/baseline_evaluation/summary.json)
evaluates the original three v1-trained models on the v1.1 inputs and targets.
It is a migration evaluation, not a new training result. The historical
three-seed training comparisons in ASSESSMENT.md and BASELINES.md still apply
to v1.0. No published-paper parity or physical accuracy gain is claimed.

## Migration evaluation of existing checkpoints

These are the original v1-trained checkpoints, evaluated once per seed on all
2,792 stage-5 test frames with the v1.1 input maps and targets. No retraining or
checkpoint selection was performed. Translation is per-axis MAE in centimetres;
rotation is mean geodesic degrees. These numbers do not measure physical
calibration accuracy.

| Training seed | Translation MAE, cm | Rotation, degrees |
| --- | ---: | ---: |
| 42 | 3.189537 | 0.729707 |
| 43 | 4.639760 | 0.813980 |
| 44 | 4.216466 | 0.782816 |
| Mean | 4.015254 | 0.775501 |

## Reproduction

Inside the public source snapshot, with the original KITTI Raw files:

```bash
python scripts/audit_calibration.py \
  --raw-root /path/to/kitti_raw --root /path/to/dataset \
  --data-output scratch/full_projection \
  --report-output scratch/calibration_audit.json
python scripts/check_calibration_release.py \
  --root /path/to/dataset --raw-root /path/to/kitti_raw
```

The audit refuses to overwrite its outputs. Compare generated Parquet hashes
with the report, or compare numerical rows if using different library versions.
To generate selected sequences directly:

```bash
kitti-c2l generate --kitti-calib-root /path/to/kitti_raw \
  --camera-model full_projection_v1_1 --output-dir scratch/frame
kitti-c2l generate-windows --kitti-calib-root /path/to/kitti_raw \
  --camera-model full_projection_v1_1 --output-dir scratch/window
```

The CLI and Python generation APIs retain `legacy_v1` as their default for
backward compatibility. Set the model explicitly for new full-projection data.

## Why no newly fitted ground truth is released

Image-edge or photometric alignment can produce a candidate refinement.
Its score can also improve by absorbing occlusion, moving objects, LiDAR spin
distortion or timestamp offsets into an incorrect rigid transform. Optimizing
and assessing the same score would not establish physical accuracy.

The current local Raw subset and released sensor assets contain image_02 and
LiDAR, but no right-camera frames, OXTS measurements or LiDAR timestamps.
The input inventory is recorded in the audit. A defensible data-driven study
would first recover those inputs or independent correspondences, model motion
and timing, fit on designated calibration/training drives, and assess stability
and accuracy on untouched drives using independent evidence. Its outputs would
be released as estimated refinements with uncertainty unless independently
validated as ground truth. That evidence has not been established here.

	# Calibration provenance and the v1.1 projection correction

	KITTI provides a carefully calibrated reference, with measurement uncertainty.
	Our v1.1 update corrects how KITTI-C2L converts that reference into a rectified
	camera-2 transform. It does not claim a new, independently measured calibration.

	## What KITTI originally calibrated

	The [KITTI data paper](https://www.cvlibs.net/publications/Geiger2013IJRR.pdf)
	describes daily calibration after recording. Camera–LiDAR registration used
	50 manually selected correspondences and a robust stereo-disparity criterion.
	It also documents the spinning LiDAR's acquisition times and camera triggering.
	Those procedures support using KITTI as a reference, but do not establish a
	sub-centimetre uncertainty bound for every driving frame.

	[Subsequent camera recalibration work](https://arxiv.org/abs/2109.03462)
	reports improvements in odometry. That is evidence that calibration can be
	refined for a particular task, rather than proof of a better camera–LiDAR
	extrinsic ground truth for KITTI-C2L.

	## Verified correction

	For a LiDAR point in metres, KITTI's documented projection is

	```text
	image_h = P_rect_02 @ R_rect_00 @ T_cam0_from_velodyne @ point_h
	```

	Writing `P = [K \| p]` gives `P = K [I \| b]`, where `b = solve(K, p)`.
	Consequently, the transform consumed by a loader that applies `K` separately is

	```text
	R = R_rect_00 @ R_cam0_from_velodyne
	t = R_rect_00 @ t_cam0_from_velodyne + solve(K, P[:,3])
	```

	The numeric source files have non-zero second and third entries in `P[:,3]`.
	Keeping only `P[0,3]/P[0,0]`, as frozen v1.0 does, discards those entries and the
	principal-point coupling in the x translation. The difference is:

	\| Capture date \| Full minus legacy translation, cm (x, y, z) \| Euclidean difference \|
	\| --- \| --- \| ---: \|
	\| 2011-09-30 \| −0.528028, −0.143972, +0.620322 \| 0.827249 cm \|
	\| 2011-10-03 \| −0.319263, −0.113119, +0.377976 \| 0.507534 cm \|

	The earlier v1 documentation's “approximately 0.62 cm” described the largest
	coordinate difference, not the full vector norm. Use the norms above when
	describing the total translation difference.

	This is an algebraic consistency correction. It does not estimate sensor poses
	from images, fit test sequences, alter raw calibration files, or show that the
	physical KITTI calibration itself was wrong by these amounts.

	## Evidence and scope

	The [machine-readable audit](../release/v1.1.0/calibration_audit.json) records
	source calibration hashes, matrices, changes and every regenerated file hash.
	The audit uses three predetermined frames per drive (10%, 50%, 90% of the
	sequence), over all ten included drives. Points are restricted to the original
	image field of view and projected depths above three metres.

	- 30 frames and 590,813 projected LiDAR points were checked.
	- Legacy conversion differs from the supplied projection by 0.22–0.50 pixels
	on average per drive, with per-drive 95th percentiles of 0.54–1.29 pixels.
	- The full conversion in float64 has a maximum discrepancy below 1e-12 pixels.
	- Actual stored float32 calibration fields have a maximum discrepancy below
	0.000039 pixels on the audited points.
	- Regeneration reproduced all four corrected Parquet files byte-for-byte.
	- All rows retain the same IDs, seeds, deltas, rotations, intrinsics, source
	mappings and splits. Only GT and miscalibrated translations change; the
	latter are recomposed, not edited independently.

	These are projection-consistency checks against the supplied KITTI matrix,
	not errors against manually observed image correspondences or independent
	physical measurements. The source projection is the reference in this test.

	The six original calibration text files are included unchanged under
	`release/v1.1.0/source_calibration/`, so source consistency can be checked
	without downloading any sensor recordings.

	The [standalone reference matrices](../release/v1.1.0/calibration_reference.json)
	provide both metre and centimetre translations, the camera intrinsics and the
	original source-file hashes.

	## Choosing a protocol

	\| Hugging Face configuration \| Calibration convention \|
	\| --- \| --- \|
	\| `metadata`, `windows` \| Frozen v1.0, x-only camera shift; unchanged default \|
	\| `metadata_full_projection`, `windows_full_projection` \| v1.1, complete numeric projection; recommended for new experiments \|
	\| `frames` \| Shared immutable KITTI sensor bytes; independent of label convention \|

	```python
	from datasets import load_dataset

	repo = "mcocheteux/KITTI-C2L-Dataset"
	metadata = load_dataset(repo, "metadata_full_projection", revision="v1.1.0")
	windows = load_dataset(repo, "windows_full_projection", revision="v1.1.0")
	```

	The existing loader materializes both conventions from their respective rows.
	Never mix GT from one convention with perturbed transforms from the other.
	Use one convention consistently for training, validation and testing, and name
	it in results. The original `v1.0.0` tag and sensor shards remain unchanged.

	The [new checkpoint evaluation](../release/v1.1.0/baseline_evaluation/summary.json)
	evaluates the original three v1-trained models on the v1.1 inputs and targets.
	It is a migration evaluation, not a new training result. The historical
	three-seed training comparisons in ASSESSMENT.md and BASELINES.md still apply
	to v1.0. No published-paper parity or physical accuracy gain is claimed.

	## Migration evaluation of existing checkpoints

	These are the original v1-trained checkpoints, evaluated once per seed on all
	2,792 stage-5 test frames with the v1.1 input maps and targets. No retraining or
	checkpoint selection was performed. Translation is per-axis MAE in centimetres;
	rotation is mean geodesic degrees. These numbers do not measure physical
	calibration accuracy.

	\| Training seed \| Translation MAE, cm \| Rotation, degrees \|
	\| --- \| ---: \| ---: \|
	\| 42 \| 3.189537 \| 0.729707 \|
	\| 43 \| 4.639760 \| 0.813980 \|
	\| 44 \| 4.216466 \| 0.782816 \|
	\| Mean \| 4.015254 \| 0.775501 \|

	## Reproduction

	Inside the public source snapshot, with the original KITTI Raw files:

	```bash
	python scripts/audit_calibration.py \
	--raw-root /path/to/kitti_raw --root /path/to/dataset \
	--data-output scratch/full_projection \
	--report-output scratch/calibration_audit.json
	python scripts/check_calibration_release.py \
	--root /path/to/dataset --raw-root /path/to/kitti_raw
	```

	The audit refuses to overwrite its outputs. Compare generated Parquet hashes
	with the report, or compare numerical rows if using different library versions.
	To generate selected sequences directly:

	```bash
	kitti-c2l generate --kitti-calib-root /path/to/kitti_raw \
	--camera-model full_projection_v1_1 --output-dir scratch/frame
	kitti-c2l generate-windows --kitti-calib-root /path/to/kitti_raw \
	--camera-model full_projection_v1_1 --output-dir scratch/window
	```

	The CLI and Python generation APIs retain `legacy_v1` as their default for
	backward compatibility. Set the model explicitly for new full-projection data.

	## Why no newly fitted ground truth is released

	Image-edge or photometric alignment can produce a candidate refinement.
	Its score can also improve by absorbing occlusion, moving objects, LiDAR spin
	distortion or timestamp offsets into an incorrect rigid transform. Optimizing
	and assessing the same score would not establish physical accuracy.

	The current local Raw subset and released sensor assets contain image_02 and
	LiDAR, but no right-camera frames, OXTS measurements or LiDAR timestamps.
	The input inventory is recorded in the audit. A defensible data-driven study
	would first recover those inputs or independent correspondences, model motion
	and timing, fit on designated calibration/training drives, and assess stability
	and accuracy on untouched drives using independent evidence. Its outputs would
	be released as estimated refinements with uncertainty unless independently
	validated as ground truth. That evidence has not been established here.