eelab-dev/Occlusion_Based_NLOS

Baseline NLOS histogram dataset

38,000 synthetic and 180 real experimental histogram samples for research on binary classification in non-line-of-sight (NLOS) sensing. Each HDF5 file contains one 32 × 32 × 64 histogram. Model weights are not included.

Contents

Source	Negative (0)	Positive (1)	Total
Synthetic	19,000	19,000	38,000
Experimental	90	90	180

Synthetic samples include clean (1,000 per class) and noise_1 through noise_36 (500 per class per condition). The condition names are identifiers, not calibrated noise magnitudes. Experimental samples retain their original acquisition groups in nine balanced folders.

data/
  synthetic_data/
    tmf8829_32x32_negative/{clean,noise_1,...,noise_36}/histogram/*.h5
    tmf8829_32x32_positive/{clean,noise_1,...,noise_36}/histogram/*.h5
  experiment_data/
    fold_1/ ... fold_9/
    manifest.csv
metadata/
  synthetic_train.csv
  synthetic_validation.csv
  synthetic_test.csv
  experimental.csv
  validation_summary.json
manifest.csv
SHA256SUMS
examples/load_sample.py
scripts/validate_dataset.py
requirements.txt

All manifest paths are relative to the repository root, including paths in data/experiment_data/manifest.csv. The root manifest indexes every sample; metadata CSVs select subsets of those same records. The Hugging Face viewer configurations expose these metadata rows, not decoded histogram tensors.

Labels and schema

label=0 means the source dataset's negative class; label=1 means its positive class. Labels are inherited from the source filenames and manifest. The exact physical criterion defining target presence/absence is not documented in the packaged source metadata and requires author confirmation.

Field	Meaning
HDF5 key	stamped_histogram
Shape	(32, 32, 64)
Axes	Two spatial pixel axes, followed by histogram bin index
Synthetic dtype	int64
Experimental dtype	uint32
Values	Stored histogram counts; no normalization applied during packaging

Spatial orientation, bin width, timing offset, and physical range units have not been verified. Do not infer them from the tensor shape. All experimental files have root attributes image_height=32, image_width=32, bin_number=64, range_min=0.0, and range_max=1890.0; the range units are unspecified.

Manifest columns: sample_id (unique path-derived identifier), path, label, label_name, source, fold, condition, split, h5_key, height, width, bins, dtype, and sha256. Synthetic fold fields are empty. Experimental condition matches the fold identifier.

Splits and experimental groups

The synthetic split assignments are preserved from the original training run: 26,600 train, 5,700 validation, and 5,700 test samples (70/15/15). The source val name is exported as validation. No new random split was generated. The source configuration specifies seed 42.

Folder	Original group	Positive	Negative
fold_1	1_experiment	10	10
fold_2	2_experiment	10	10
fold_3	3_experiment	10	10
fold_4	4_experiment	10	10
fold_5	5_experiment	10	10
fold_6	addition_test/left	10	10
fold_7	addition_test/right	10	10
fold_8	addition_test/long	10	10
fold_9	addition_test/short	10	10

Experimental samples have split=unassigned: these folders preserve acquisition groups and do not themselves specify a cross-validation protocol. For evaluation, state whether real data are held out entirely or used in group-based cross-validation. Keep each acquisition group together when constructing a new experimental split. The historical use of these samples for model selection has not been audited.

The supplied synthetic split is a historical baseline, not a guarantee of independence across related clean/noisy scenes. Scene-family identifiers are not included, so scene-level leakage cannot be ruled out. Fit preprocessing statistics on training data only and report the split and preprocessing used.

Loading a sample

python -m pip install -r requirements.txt
python examples/load_sample.py --source experimental --fold fold_1
python examples/load_sample.py --source synthetic --split train

After publication, replace PeizhaoEdinburgh/Occlusion_Based_NLOS with the repository ID:

python examples/load_sample.py --repo-id PeizhaoEdinburgh/Occlusion_Based_NLOS --source experimental --fold fold_1

Use --revision COMMIT_HASH to pin the remote version. The example downloads the manifest and one sample, verifies its SHA-256 hash and schema, and loads it with h5py. It returns the stored values without training-specific preprocessing. For direct access:

import h5py

with h5py.File("data/experiment_data/fold_1/1_positive_histogram_capture_1.h5", "r") as f:
    histogram = f["stamped_histogram"][()]  # shape (32, 32, 64)

Provenance and limitations

The source collection is identified as LameDar2/tmf8829_32x32 and the baseline run as 20260726_212202_tmf8829_36noise_seed_lhs_plus30_36_global_log_count_coarse_multiscale_h128_d4_aug_standard_hist_coarse_multiscale_mlp. The synthetic directory names identify the TMF8829 sensor setting. Complete simulation parameters, sensor calibration, acquisition dates, physical setup measurements, and noise-condition definitions are not supplied here.

The experimental files were moved into the fold structure without changing their contents. Two spelling errors in filenames (positve, negtive) were corrected. Original experiment group names are recorded above. Machine-specific source paths are excluded from the portable manifests.

This dataset is intended for research. Balanced class counts do not establish real-world prevalence or generalization to other sensors and environments. No paper citation or author attribution has been supplied yet.

Integrity

python scripts/validate_dataset.py
# Alternatively, verify file bytes only from the repository root:
sha256sum --check SHA256SUMS

The validator checks all 38,180 sample hashes and HDF5 schemas, complete manifest coverage, split consistency, and the nine balanced experimental groups.

License and publication status

The repository specifies the MIT license. Author confirmation is still needed for attribution, label semantics, physical units, and acquisition/generation details.

Dataset repository: PeizhaoEdinburgh/Occlusion_Based_NLOS. Local runtime and environment directories such as .omx/ and .venv/ are excluded from the published package.

The metadata layout follows Hugging Face's manual configuration documentation. Remote loading uses the documented Hub download API.