Door Identification -- Experiments

Eyal Amir and Pedrito Maynard-Reid II

Our experimental results are shown below. The first column includes the input images (size reduced by a factor of 4) and the second column includes the output images with crosses indicating detected door centers. The lines superimposed on the image indicate the boundary lines of the door. (Some superimposed lines appear segmented because of the image reduction; clicking on any image will produce a full-sized version with all the lines correctly displayed.) We chose not to discard "noise" lines so as to allow for the examination of the procedure results. To reproduce the experiments, take a look at the code mentioned in the Appendix:code section and in the file experiments.m.

The first set contains routine door fronts and corridors.

Original ImageOutput ImageCommentary
lecture202b.gif: The easiest case -- one door, straight ahead, almost no similarly colored objects.

Difficulties: different lighting conditions for the upper and lower parts of the door.

Difficulties: different lighting conditions for the upper and lower parts of the door. The unlit part of the carpet has the same color as the bottom of the door.
open214.gif: Continues to work when door is partially open.
corridor196.gif: No visible doors, none detected
2a2corridor201b.gif: Found all and only the doors, including the door segment in the lower left corner.
hall208.gif: Found doors near and far, straight ahead and on the side, all in the same image. The door sliver in the middle was not wide enough to pass the width threshold.
hall200b.gif: The corkboards in this and the following two images look similar enough to small doors to get detected.
corridor2A215.gif: Missed far door because person occluded one edge.
corridor206.gif: Missed near left door because of excessive reflection.
corridor203b.gif: Missed right door because distortion slanted the sides too much. Split bathroom door in two because of occlusion by sign.

The second set features trickier settings such as lobbies and the insides of sunlit offices. The algorithm was much less successful in these situations.

Original ImageOutput ImageCommentary
robotics210b.gif: Found many strange lines because of the reflection and interreflection caused by a different light source type and location than we designed for.
ksl199.gif: An extreme case of false positives -- found all the doors, but also found many vertical, door-colored objects such as parts of the sofa frame and the corkboard.
library204.gif: Found doors using only frames around glass. Interestingly, door-colored objects seen through the glass also got detected. The color of the shaded white wall was close enough to the door brown to be classified.
window217.gif: Another example where heavy shading caused non-brown pieces of objects to get classified as doors.
open218.gif: Occlusion from the lamp made the side of the desk appear to be two objects.
eyal216.gif: Missed the left door edge because of severe reflection from the sun through the office window.

Next: Conclusion

Return to Table of Contents

Eyal Amir and Pedrito Maynard-Reid II.
Last modified: Tue Mar 16 19:25:17 PST 1999