Geometric Size Estimation from Camera View

This project demonstrates a method to estimate real-world object dimensions from a first-person camera perspective using ray-plane intersections and camera modeling. It features:

• Utility functions for estimating real-world dimensions from screen-space coordinates
• A visual proof of concept for intuitive understanding
• A command-line interface for estimating object sizes from user input
• Unit tests to verify the implementation

The core folder contains all the logic related to the geometric size estimation, while the rest of the project is dedicated to providing utility/test scripts to use it. Most user should only use the cli folder to estimate object sizes from the command line or the fps_demo folder to visualize and interact in an fps-like environment (instructions below). Advanced user might create custom scripts that leverage functions from the core folder.

---

Examples

fps_demo/viewer.py

Run with:

python -m fps_demo.viewer

Controls:

• QZSD — Move the box (left / up / down / right)
• WXCV — Change the box size
• RF — Change the camera height
• TGYH — Change the camera field of view
  (T / G for horizontal FOV, Y / H for vertical FOV)
• Mouse — Control camera yaw and pitch

---

cli/estimate.py

Example usage:

python -m cli.estimate --image-width 768 --image-height 432 --bbox-x 300 --bbox-y 100 --bbox-width 150 --bbox-height 100 --focal-length-mm 4.6 --sensor-width-mm 5.6 --sensor-height-mm 3.2 --camera-height 2.12 --yaw 0.0 --pitch -25.5

Result:

Estimated physical size of object:
Width  = 1.493 m
Height = 1.009 m

Assumptions:
- Camera position: [0.   2.12 0.  ]
- Yaw: 0.00 deg, Pitch: -25.50 deg
- Resolution: 768 x 432
- Focal length: 4.6 mm
- Sensor size: 5.6 mm x 3.2 mm
- Box coordinates: (300.0, 100.0), Width: 150.0, Height: 100.0
- Box spanning from (300.0, 100.0) to (450.0, 200.0)
- Camera FOV: 60.0 deg horizontal, 33.75 deg vertical

Estimates the size of an object given user input (command-line arguments).

---

Tests

• fps_demo/viewer.py: Manual visual test to verify whether the results make intuitive sense.
• tests/test_plane.py: Verify the Plane class.
• tests/test_camera.py: Verify the Camera class.
• tests/test_intersection.py: Verify the intersection logic.
• tests/test_largest_diameters.py: Verify the largest diameters computations.
• tests/test_against_YOLO.py: Compare the estimated sizes against the data/yolo_groundtruth_data.csv dataset and verify the results do not diverge too much.

test ideas (TODO):

• quantify the error if we add lens distortion

---

Current Limitations

• Assumes a flat ground plane at y = 0.
• Uses a simplified pinhole camera model (no lens distortion).
• The object size is estimated by adding the distance between the intersection of the box's center ray with the ground plane and each 4 plane defined by the bounding box corners and the camera's parameters. The estimated width is the sum of the left plane-intersection and intersection-right plane distances. The estimated height is the sum of the top plane-intersection and intersection-bottom plane distances. The accuracy of this method depends on the real-world context. In the case of a floating object, it makes sense to assume the center of the object is at the ground level.
• Camera roll is not yet supported, but could be implemented with minimal changes.

Citation

If you use this code, please cite:

@article{grimmer2025debris,
title = {A geometric and deep learning reproducible pipeline for monitoring floating anthropogenic debris in urban rivers using in situ camera},
journal = {Submitted},
volume = {},
pages = {},
year = {2025},
issn = {},
doi = {},
url = {},
author = {Gauthier Grimmer and Romain Wenger and Clément Flint and Germain Forestier and Gilles Rixhon and Valentin Chardon}
}