xuguanyu04/PHAROS-data

PHAROS data

Recordings from the PHAROS soft optical waveguide sensor, used by the code in PHAROS. For the indentation data, every optical frame is paired with a depth image of the deformed membrane from a RealSense camera, which the models use as ground truth.

Folder	Content
indentation/	Paired optical rows and depth images of the membrane pressed by different indenters
bending/	Photodiode logs recorded at different bending angles, as CSV
stretch/	Optical rows and depth images of a stretching recording
checkpoints/	Trained model used in the paper (checkpoints/final-run/)

indentation/
  white_train_S1.npz ... white_train_S7.npz                     training sessions
  white_test_Circle.npz, _Finger, _Square, _Triangle, _U.npz    held-out test sessions
bending/   34 CSV files
stretch/   stretch.npz
checkpoints/final-run/   trained model and evaluation results

Download

pip install -U huggingface_hub
hf download xuguanyu04/PHAROS-data --repo-type dataset --exclude "checkpoints/*" --local-dir data      # datasets
hf download xuguanyu04/PHAROS-data --repo-type dataset --include "checkpoints/*" --local-dir .         # trained model

indentation/

Each .npz archive is one packed session:

Array	Shape	Type	Description
depth	(N, 230, 230)	uint16	Depth crop around the membrane, in units of meta_depth_scale_m_per_unit (0.0005 m)
optical	(N, 218)	int32	Raw optical row: time_ms, then L{i}P{j} for LED i in 0..30 and photodiode j in 0..6
frame_ids	(N,)	int64	0 to N-1, in acquisition order
mtime_unix	(N,)	float64	Modification time of the recorded frame
depth_unit_scale	(N,)	float32	2.0 for frames recorded in half the nominal depth unit, else 1.0
meta_*			Camera intrinsics, region of interest and capture settings (see below)

Index 0 of the LED and photodiode axes is the "off" baseline. The 180 channels used as model input are LEDs 1 to 30 times photodiodes 1 to 6.

Split	Sessions	Frames
Train	white_train_S1 to white_train_S7	345,571
Test	white_test_Circle (817), _Finger (851), _Square (1066), _Triangle (879), _U (899)	4,512

bending/

One CSV per recording, named ADPD_2211_test_<angle>deg_<haslight|nolight>_<date>_<time>.csv. The angles run from -210 to 210 degrees in steps of 30, and each angle was recorded with the LEDs on (haslight) and off (nolight). Columns:

TIME, EXT, L1P1, L1P4, L2P1, L2P4, L3P1, L3P4, P1O, P4O

L<i>P<j> is the reading of photodiode j while LED i is on.

stretch/

stretch.npz holds four arrays:

Array	Shape	Type
depth	(81467, 230, 230)	uint16
optical	(81464, 218)	int32
depth_frames	(81467,)	int32
optical_frames	(81464,)	int32

The depth and optical recordings are not row-aligned. depth_frames[i] and optical_frames[i] give the frame number of row i in each array. 81,359 frame numbers occur in both, 108 only in depth and 105 only in optical. Pair the rows by frame number:

checkpoints/

checkpoints/final-run/ is the model reported in the paper, trained on the indentation training sessions:

Path	Content
best_combined.pth	Final model: photodiode readings (180 channels) to a 4096-point cloud. Latent size 256, input min-max normalised with constants stored in the file
checkpoints/best_model.pth, checkpoints/last.pth	Point-cloud autoencoder (best validation loss, and the last step)
eval_ae_test/	Autoencoder reconstruction metrics and example figures on the test sessions
eval_sessions/	Per-session metrics of the full model on the test sessions

Load and evaluate it with the code repository:

hf download xuguanyu04/PHAROS-data --repo-type dataset --include "checkpoints/*" --local-dir .
python scripts/evaluate.py --run-name final-run
python scripts/evaluate_sessions.py --run-name final-run