Home People Research Publications Demos
News Jobs Prospective
About Internal

Semantic image classification results

Below we illustrate semantic classification results obtained on the Corel database, and a database of personal photographs assembled by Kodak, and hand-labeled for concepts such as indoor/outdoor, sky, grass, etc. More details can be found in the ICIP'02 paper [ps][pdf].

Corel - 15 semantic classes

Kodak - Results for sky detection and indoor/outdoor classification

Example errors

Frequently one learns more about the goodness of a classifier by inspecting the errors than from statistics such as those above. Below are examples of typical errors on Kodak.

18.5% of the indoors images classified as outdoors

These errors tend to be either 1) images that contain significant amounts of outdoors (e.g. large windows) or 2) a large ground plane.

15.6% of the outdoors images classified as indoors

These errors tend to images that would require a substantial amount of "high-level" knowledge about scenes, people, houses, etc. to classify correctly (e.g. "porches look like living rooms but they have a wooden deck instead of a wooden floor", "porch chairs are different than living room chairs", and so on).

15.6% of the sky false positives

These errors tend to be images that have large smooth areas with colors that are similar to those of sky or clouds.

29.4% of the sky misses

These errors tend to be images that have very small regions of unoccluded  sky.
Contact: Nuno Vasconcelos