ngocleltt/lamp-triangulation

Lamp Triangulation Dataset

This repository contains the public release of my 2nd Dataset from the lamp triangulation project. It focuses on a short road segment along Malyy Moskvoretskiy Bridge and ul. Bolshaya Ordynka, covering 15 street lamps that were manually identified, tracked, and triangulated from dashcam footage.

Repository structure

lamp-triangulation/
├── README.md
├── LICENSE
├── assets/
│   ├── lamp_map_polyline_paper_clean.png
│   └── lamp_quality_rank.png
├── scripts/
│   ├── azimu-elevation.py
│   ├── lamp_id_tracking.py
│   ├── make_lamp_map_2d.py
│   ├── make_lamp_quality_assessment.py
│   └── triangulate_lamps.py
└── data/
    ├── annotations/
    ├── images/
    ├── labels/
    ├── classes.txt
    ├── lamp_quality_summary.csv
    ├── metadata.csv
    └── triangulation_multiview_ready.csv

The dataset was prepared for a pipeline that goes from metadata and tracking to center extraction, angle estimation, multiview filtering, triangulation, quality assessment, 2D mapping, and final dataset selection for paper use.

Main tabular schema

The exact columns in triangulation_multiview_ready.csv reflect the reconstruction pipeline, but the file is centered around the following information:

• lamp_id: identifier of the lamp.
• file_name: image or frame name containing the lamp.
• frame_number: frame number in the video or route sequence.
• frame_index: sequential index in the metadata.
• timestamp_sec: time of the frame in seconds.
• latitude, longitude: GPS position of the vehicle at the moment the frame was captured.
• source: data source, here dashcam.
• road_segment: route segment label.
• num_objects: total number of objects in the frame.
• num_lamp_on, num_lamp_off, num_lamp_occluded: lamp state counts.
• active_lamps: lamp IDs visible or active in the frame.
• assigned_track_ids: tracked lamp IDs assigned to the frame.
• note: optional note field.
• bbox_center_x_norm, bbox_center_y_norm, bbox_width_norm, bbox_height_norm: normalized bounding box values.
• bbox_center_x_px, bbox_center_y_px: bounding box center in pixels.
• label_match_status: label matching status such as single_box, matched_by_lamp_id_order, or fallback_first_box.
• camera_cx, camera_cy: image center in pixels.
• fx_px, fy_px: focal length in pixels.
• azimuth_deg, elevation_deg: lamp viewing angles relative to the camera center.
• num_views_for_lamp: number of valid observations for the lamp across the dataset.
• is_multiview_ready: flag indicating whether the lamp has enough valid views for triangulation.

The dataset is therefore suitable for both annotation analysis and triangulation-based localization workflows.

Pipeline overview

The project pipeline is organized as follows:

1. Metadata preparation.
2. Lamp tracking and ID assignment.
3. Center extraction.
4. Azimuth and elevation estimation.
5. Multiview filtering.
6. Triangulation prototype in 3D.
7. Quality assessment.
8. 2D map reconstruction.
9. Heatmap and ranking analysis. This release corresponds to the post-triangulation, pre-final-analysis stage, where the reconstruction outputs and quality metrics are already available.

Suggested use cases

• Lamp triangulation experiments.
• Reconstruction and localization benchmarks.
• Quality inspection of triangulated lamp candidates.
• Annotation and metadata analysis.
• Applied computer vision research on street-lamp mapping.