ngqtrung/ptl-crossing-check

PTL Crossing Check

Street-crossing photos from the pedestrian point of view, for one narrow decision for a blind walker at the kerb: GO (the pedestrian light is steady green) or WAIT. Each photo also has the direction of the crossing and a flag for people or vehicles on the crossing.

This is a re-packaged copy of the Pedestrian-Traffic-Lights (PTL) dataset from the ImVisible project (Samuel Yu, Heon Lee, John Kim / Jung Hoon Kim; photos from Shanghai; GitHub, MIT licence), in the 876x657 resolution. The images, the five signal classes, the zebra endpoints and the blocked tag are the original human labels from the PTL authors. The splits are the original PTL splits (so results compare with the published LytNet numbers). We added only the derived columns that the table below marks as "derived". We did not relabel any image.

split	images	red	green	countdown_green	countdown_blank	none	blocked
train	3456	1016	879	648	621	292	742
validation	864	235	239	171	153	66	191
test	739	226	185	144	130	54	145

Columns (metadata.csv in each split folder)

column	source	meaning
file_name	original	image file (876x657 JPEG)
signal_id	original	PTL class: 0 red, 1 green, 2 countdown green, 3 countdown blank, 4 none
signal	derived (rename)	red, green, countdown_green, countdown_blank, none
go	derived (rule)	True only when signal == green
x1, y1, x2, y2	original	the two endpoints of the zebra-crossing midline, in pixels of the original 4032x3024 photo. In 5058 of 5059 images, (x2, y2) is the lower point (near end, at the walker) and (x1, y1) is the upper point (far end)
x1_norm … y2_norm	derived (rule)	the same points divided by 4032 (x) and 3024 (y), so they are valid for any resolution
heading_deg	derived (rule)	angle of the midline from the near end (x2, y2) to the far end (x1, y1): degrees(atan2(x1 - x2, y2 - y1)); 0 = straight up the image, negative = the crossing goes to the left, positive = to the right
heading	derived (rule)	left if heading_deg < -15, right if heading_deg > 15, else ahead
blocked	original (to bool)	the PTL block tag: True = blocked, False = not_blocked

How the derived labels were made

All derived columns come from a fixed rule on the original labels. No model and no person made new labels for them.

• go: the safety rule of the use case. Countdown green also has a green light, but a blind walker cannot read the number of seconds, so the app does not start a crossing on a countdown. Countdown blank, red and none are WAIT.
• heading: the ±15° band is our choice for a 3-way spoken cue ("crossing ahead / to your left / to your right"). Use heading_deg if you want a different band. Check the sign convention before you use the cue: the value is the angle as seen in the image, not a compass bearing.

Code: poc/crossing/export_hf.py and poc/crossing/metrics.py (function labels) in our project.

Labels we made that are NOT in this dataset

For an out-of-distribution test, we labelled 105 frames from GuideDog (walking videos from many countries) for the pedestrian signal (walk, walk_countdown, dont_walk, dont_walk_countdown, none, unclear) and the walker position (at_kerb, on_crossing, near, no). Two independent Claude Sonnet passes labelled each frame with a written guide. We kept a frame only when both passes agreed. These are model labels, not human labels. GuideDog is CC BY-NC 4.0 and gated, so we do not redistribute those frames or their labels here.

Use case and results (test split, 739 images)

The walker hears GO only when p(green) ≥ τ. τ is the lowest value with an unsafe-GO rate (GO on red or none) of 1 % or less on validation; the same τ is applied to test.

model	signal acc	macro-F1	heading acc	blocked AUC	latency p50
d3-flash (8.4B) + LoRA, our fine-tune	0.978	0.976	0.877	0.953	701 ms (H100)
d3 26B, zero-shot	0.708	0.597	0.206	0.864	1143 ms (H100)
LytNet V1 (published weights)	0.938	0.924	0.870	–	2 ms
LytNet V2 (published weights)	0.957	0.948	0.824	–	3 ms

At its validation τ, the fine-tuned d3-flash has unsafe-GO 0.7 % and GO recall 99.5 % on test. LytNet V2 has 0.7 % and 90.3 %. The latencies were measured while the GPU was shared, so they are approximate.

Limit. All photos come from one city (Shanghai). On the 105 GuideDog frames from other countries, the fine-tuned model gave a wrong GO on 8 of 74 dangerous frames. Do not use a model trained on this data as a safety device.

Licence and citation

The PTL data is from the ImVisible repository, which is under the MIT licence. Keep the copyright notice of the original authors. Please cite their papers:

@article{yu2019lytnet,
  title   = {LYTNet: A Convolutional Neural Network for Real-Time Pedestrian Traffic Lights and Zebra Crossing Recognition for the Visually Impaired},
  author  = {Yu, Samuel and Lee, Heon and Kim, John},
  journal = {arXiv preprint arXiv:1907.09706},
  year    = {2019}
}
@article{yu2019lytnetv2,
  title   = {Street Crossing Aid Using Light-weight CNNs for the Visually Impaired},
  author  = {Yu, Samuel and Lee, Heon and Kim, Jung Hoon},
  journal = {arXiv preprint arXiv:1909.09598},
  year    = {2019}
}