LOVEIII486/isy5002-lta-traffic-images

LTA Traffic Camera Images with Human-Corrected Vehicle Boxes (Singapore, September 2026)

Two weeks of stills from Land Transport Authority (LTA) traffic cameras in Singapore, September 2026, together with a small human-corrected detection subset.

15,316 images · 8 cameras · 2.6 GB · 10-minute sampling

63 human-corrected images · 1,197 vehicle boxes · 8 cameras

Attribution

Contains information from the Traffic Images dataset accessed on 2026-09-13 to 2026-09-29 from data.gov.sg which is made available under the terms of the Singapore Open Data Licence version 1.0

This dataset is not endorsed by, affiliated with, or an official product of the Singapore Government or the Land Transport Authority. The human-corrected annotations were produced by the project team and are not an LTA product.

Contents

Directory	Cameras	Images	Period (UTC)
week1/	2701 2702 2704 4703 4712 4713 4798 4799	7,321	2026-09-13 → 2026-09-20
week2/	the same 8	7,995	2026-09-22 → 2026-09-29
annotations/lta_human_v1/	the same 8	63 annotated images	week 2

week1/2701/20260913T124023Z_2701_1240_97acd1b15491.jpg
└──┬─┘└┬─┘ └─────┬──────┘ └┬─┘ └┬─┘ └────┬─────┘
  week camera  captured_at  camera HHMM  sha256[:12]
                 (UTC)          (Singapore time)

Human-corrected vehicle annotations

annotations/lta_human_v1/ contains a 63-image human-corrected detection subset. It is not a full annotation of all 15,316 images: it covers 63 of the 80 week-2 validation frames used by the project (78.8% of that validation set, or 0.41% of the full collection). The other 17 validation frames were removed because the annotator judged them too dense or too dark to label reliably.

What was annotated

One anonymous annotator reviewed the 80 week-2 validation frames in Label Studio. Each frame started from YOLO26x teacher pre-labels generated at the original 1920-pixel resolution (confidence threshold 0.3). The annotator checked every vehicle, corrected classes and box extents, added missed vehicles, and removed boxes that were not vehicles. The labels are therefore human-corrected teacher pre-labels, not an annotation from scratch. This package contains detection boxes only; the project's separate 600-frame congestion-level labels are not included here.

• 63 images retained, 17 removed
• 36 daytime and 27 nighttime images
• 1,197 boxes: 611 cars, 459 motorcycles, 30 buses and 97 trucks
• 96% of the teacher boxes retained
• 566 boxes added, including 385 drawn by hand
• One anonymous annotator; no inter-annotator agreement

Camera	Images
2701	7
2702	3
2704	8
4703	10
4712	10
4713	8
4798	8
4799	9
Total	63

The 17 excluded frames are concentrated on camera 2702 and at night. Metrics on this subset are therefore likely optimistic for dense and nighttime scenes, and the subset is a human-corrected validation subset, not an independent test set: the same images were used for teacher-reference comparisons and resolution selection, and the teacher family has an advantage because the annotations start from its boxes.

Annotation files

Path	Description
images.csv	63 rows: image path, camera, day/night, UTC time, dimensions, SHA-256, box count
split.csv	The same rows in the evaluator's split format
val.txt	Image paths relative to the dataset root (week2/...)
labels/val/*.txt	YOLO labels (class cx cy w h, normalized)
annotations/instances.json	COCO detection annotations in absolute pixels
annotations/annotations.jsonl	One JSON object per image, convenient for datasets
excluded.csv	The 17 removed images and the reason
dataset.json	Provenance, class names, counts and label-manifest hashes
checksums.sha256	SHA-256 for every annotation-package file

Images are not duplicated. Join each image_path in images.csv to the corresponding file at the dataset root, and verify it against the recorded SHA-256. The YOLO class order is:

0 car   1 motorcycle   2 bus   3 truck

The JSONL file can be loaded directly:

from datasets import load_dataset

annotations = load_dataset(
    "json",
    data_files="annotations/lta_human_v1/annotations/annotations.jsonl",
    split="train",
)
print(annotations[0])

Source and collection

Images were fetched from the data.gov.sg real-time traffic images API by a small scheduled script run across roughly 30 collection windows. Every round requested all 8 cameras; a frame was written only when it was fresh and its sha256 differed from the previous one, so the archive contains no duplicate frames.

Camera feeds freeze overnight — the longest observed gap was 215 minutes — so the freshness limit was raised from 15 to 240 minutes on 2026-09-16. The resulting bias is that 00:00–07:00 on 2026-09-14 and 09-15 is markedly under-represented.

The human-corrected annotations were added by the project team. The source images are the week-2 validation frames; the annotation generation and filtering commands are in src/detection/annotate.py in the project repository.

Intended use

Vehicle detection and counting, traffic density estimation, diurnal pattern analysis, per-camera time-series modelling, and evaluation on the human-corrected 63-image subset.

Limitations

• The base images have no annotations. A 63-image human-corrected detection subset is provided under annotations/lta_human_v1/; it is not a full annotation of all images.
• The annotation subset is teacher-assisted and small. It was initialized from YOLO26x pre-labels, produced by one anonymous annotator, and covers only 63 images. It is a validation subset, not an independent test set; see annotations/lta_human_v1/README.md.
• Night frames are dark, and detection recall drops sharply.
• No geographic metadata, only camera IDs; lane geometry must be derived from the images if needed.
• Number plates may be legible in some full-resolution frames. The Singapore Open Data Licence does not extend to personal data; consider this before redistributing further.

License

The images remain under the Singapore Open Data Licence 1.0. The human-corrected annotation files are released under CC BY 4.0. Attribution: "LTA human-corrected vehicle boxes, ISY5002 Project Team (anonymous), CC BY 4.0." See annotations/lta_human_v1/dataset.json for the machine-readable record.