Real-time efficient perception is critical for autonomous navigation and city scale sensing. Orthogonal to architectural improvements, streaming perception approaches have exploited adaptive sampling improving real-time detection performance. In this work, we propose a learnable geometry-guided prior that incorporates rough geometry of the 3D scene (a ground plane and a plane above) to resample the images for efficient object detection. This significantly improves small and far-away object detection performance while also being dramatically more efficient both in terms of latency and memory. For autonomous navigation, using the same detector and scale, our approach improves detection rate by +4.1 APS or +39% and in real-time performance by +5.3 APS or +63% for small objects over state-of-the-art. For fixed traffic cameras, our approach detects small objects at image scales other methods cannot. At the same scale, our approach improves detection of small objects by +195% or +12.5 APS over naive-downsampling and +63% or +4.2 APS over state-of-the-art.

Committee:
Prof. Srinivasa G. Narasimhan (co-advisor)
Dr. Christoph Mertz (co-advisor)

Prof. Deva Ramanan

Gaurav Parmar