ngocleltt/lamp-state-detection

Lamp State Detection Dataset

A manually annotated computer vision dataset for street-lamp bulb detection from road-view images, designed for research on lamp-state recognition and downstream geo-aware infrastructure monitoring.

This dataset focuses on the lamp bulb itself, not the entire lamp post. Each image is annotated with bounding boxes for visible bulbs and labeled by state. The current label set includes lamp_on, lamp_off, and lamp_occluded. The dataset also includes metadata such as timestamp and geographic coordinates for each image sample.

This dataset contains 8428 (updated 16 June) manually annotated JPG images with JSON files, YOLO TXT files, and a metadata CSV file. The images are all 1280 × 720 pixels, and the label set includes 3 classes: lamp_on, lamp_off, and lamp_occluded.

This dataset is intended as the first stage of a broader applied pipeline:

• Detect the lamp bulb and classify its visible state.
• Associate detections with image metadata such as time and GPS.
• Support future localization, mapping, and triangulation workflows for faulty street lamps.

---

Dataset contents

The dataset currently contains the following components:

• images/: JPG images.
• annotations/: JSON annotation files.
• labels/: TXT label files in YOLO format.
• metadata.csv: per-image metadata with file path, video/frame identifiers, timestamp, geographic coordinates, label summary, number of objects, bounding box summary, and image size.

Route coverage

This dataset was collected in Moscow across three main tram routes:

1. A highway route from the Kuskovo area toward the city center.
2. A route around Moscow City, starting near Kutuzovsky Ave and continuing along the A106 toward Rublyovo-Uspenskoye.
3. A Ramenki route heading toward Aminevskoye / Rublevskoye.

Metadata fields include:

• relative_path
• file_name
• video_id
• frame_id
• timestamp_sec
• latitude
• longitude
• location_source
• location_method
• labels_present
• num_objects
• bboxes_summary
• image_width
• image_height

---

Why dataset matters? Most publicly available street-light datasets and demo models emphasize detecting the whole streetlight pole or fixture, while this dataset is built around the finer-grained target of the street-lamp bulb region. That distinction matters for practical monitoring tasks where the operational state of the light source itself is more important than detecting the pole structure.

Task definition

The current release supports object detection for three visual classes:

1. lamp_on: the bulb is visibly illuminated.
2. lamp_off: the bulb is visible but not illuminated.
3. lamp_occluded: the bulb region is partially blocked or visually ambiguous because of occlusion.

The target object is the bulb/light-emitting region, not the full pole, support arm, or surrounding infrastructure. This is the main conceptual difference between this dataset and many broader streetlight detection datasets.

Directory structure

lamp-state-detection/
├── images/
├── annotations/
├── labels/
├── metadata.csv
├── data.yaml
└── README.md

Limitations

This release is primarily a 2D detection dataset with supporting metadata. It is not yet a full triangulation dataset, because it does not by itself define multi-view correspondence, camera pose estimation, or final ground-truth 3D lamp coordinates for each object instance. Users should therefore treat this release as a perception-oriented dataset and as the first layer of a larger applied system for street-lamp monitoring.

Acknowledgment

This dataset was created as part of an applied research direction on street-lamp fault detection, geo-referenced monitoring, and future triangulation-based localization workflows.