eduzhai > Applied Sciences > Engineering >

Pixel Invisibility Detecting Objects Invisible in Color Images

  • Save

... pages left unread,continue reading

Document pages: 11 pages

Abstract: Despite recent success of object detectors using deep neural networks, theirdeployment on safety-critical applications such as self-driving cars remainsquestionable. This is partly due to the absence of reliable estimation fordetectors failure under operational conditions such as night, fog, dusk, dawnand glare. Such unquantifiable failures could lead to safety violations. Inorder to solve this problem, we created an algorithm that predicts apixel-level invisibility map for color images that does not require manuallabeling - that computes the probability that a pixel region contains objectsthat are invisible in color domain, during various lighting conditions such asday, night and fog. We propose a novel use of cross modal knowledgedistillation from color to infra-red domain using weakly-aligned image pairsfrom the day and construct indicators for the pixel-level invisibility based onthe distances of their intermediate-level features. Quantitative experimentsshow the great performance of our pixel-level invisibility mask and also theeffectiveness of distilled mid-level features on object detection in infra-redimagery.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×