Abstract: Over the last few decades, many architectures have been developed that
harness the power of neural networks to detect objects in near real-time.
Training such systems requires substantial time across multiple GPUs and
massive labeled training datasets. Although the goal of these systems is
generalizability, they are often impractical in real-life applications due to
flexibility, robustness, or speed issues. This paper proposes RMOPP: A robust
multi-objective post-processing algorithm to boost the performance of fast
pre-trained object detectors with a negligible impact on their speed.
Specifically, RMOPP is a statistically driven, post-processing algorithm that
allows for simultaneous optimization of precision and recall. A unique feature
of RMOPP is the Pareto frontier that identifies dominant possible
post-processed detectors to optimize for both precision and recall. RMOPP
explores the full potential of a pre-trained object detector and is deployable
for near real-time predictions. We also provide a compelling test case on
YOLOv2 using the MS-COCO dataset.