EPrOD: Evolved Probabilistic Object Detector with Diverse Samples

Jaewoong Choi, Sungwook Lee, Seunghyun Lee, Byung Cheol Song

Abstract: Even small errors in object detection algorithms can lead to serious accidents in critical fields such as factories and autonomous vehicles. Thus, a so-called probabilistic object detector (PrOD) has been proposed. However, the PrOD still has problems of underestimating the uncertainty of results and causing biased approximation due to limited information. To solve the above-mentioned problems, this paper proposes a new PrOD composed of four key techniques, i.e., albedo extraction, soft-DropBlock, stacked NMS, and inter-frame processing. In terms of Probability-based Detection Quality (PDQ) metric, the proposed object detector achieved 22.46, which is 4.46 higher than a backbone method, for the Australian Centre for Robotic Vision (ACRV) validation dataset consisting of four videos.