Depth of Field: Optical Equivalence, Computational Bokeh and Failure Cases

Depth of field is governed by focal length, aperture, focus distance, sensor format and the acceptable circle of confusion for the final image. Computational bokeh estimates depth and applies simulated blur; it does not reproduce every optical effect of a larger aperture or a different sensor format. A fair comparison must control framing, viewpoint, subject distance and output size.

For a technical explanation of the circle of confusion and the optical basis of depth of field, see MIT CSAIL’s computational photography explanation of depth of field and circle of confusion.

For a related comparison of computational blur and optical depth cues, see our guide to computational bokeh.

Define optical equivalence

To compare two camera systems, use the same camera position, field of view, subject distance and final framing. Record sensor dimensions, focal length, aperture, focus distance and output size. If the camera position or framing changes, perspective and depth relationships may change too. Circle-of-confusion assumptions should be tied to the final viewing size and distance.

For the system-level differences between full-frame and medium-format capture, see full frame versus medium format.

Optical depth of field

Aperture controls the cone of light reaching the sensor and influences depth of field and diffraction. Focal length and focus distance also affect the range that appears acceptably sharp. Background blur is not only blur amount: lens aberrations, pupil shape, vignetting and highlight rendering affect the appearance of bokeh.

For the optical design factors that influence image detail and rendering, see aspherical lens technology.

Computational depth estimation

Portrait modes estimate a subject mask or depth map, then apply blur based on the estimated distance from the focus plane. The depth map may come from multiple cameras, dual-pixel data or learned inference. The software must infer boundaries and occlusions that a physical lens renders optically.

Failure examples to test

  • Hair against a bright background: missing strands or halo edges.
  • Glass or transparent objects: incorrect foreground/background classification.
  • Fences and fine mesh: holes filled or background incorrectly blurred.
  • Overlapping limbs: depth ordering errors at occlusion boundaries.
  • Moving subjects: inconsistent segmentation between frames.
  • Point lights: synthetic blur discs that do not match optical highlights.

Controlled comparison protocol

  1. Use a named smartphone and a named interchangeable-lens camera.
  2. Match viewpoint, framing, subject distance and output dimensions.
  3. Capture a natural optical-depth-of-field reference and computational portrait output.
  4. Compare edge accuracy, transition quality, foreground blur, highlight rendering and movement artefacts.
  5. Repeat at multiple subject distances and lighting conditions.
  6. Preserve source files and record device/software versions.

Conclusion

Optical and computational depth-of-field effects can produce similar-looking images in favourable conditions, but they are generated by different processes. Equivalence requires controlled framing and viewing conditions, and the most informative evaluation includes difficult segmentation and depth-ordering cases.