Cascaded inference for high quality single-shot object detection

Abstract

In object detection, most accurate detectors are based on two-stage detectors based on R-CNN approach, where candidates are proposed and further refined after feature pooling step in cascaded manner. In contrast, one-stage detectors have advantages of being more efficient and simpler, but have lagged the accuracy of two-stage thus far. In this paper, we argue that one-stage detectors mainly suffer from architectural limitation, i.e, missing two-step cascaded inference mechanism that is applied in twostage detectors. Due to the lack of this mechanism, one-stage detectors struggle with two main arising issues: 1) sensitiveness of pre-defined anchor configuration, due to current heuristic anchor matching strategy, and 2) misalignment of receptive field in detection head induced by the fixed receptive field of the standard convolutional detection head. To incorporate the cascaded inference mechanism into the one-stage detectors and thus overcome these issues, we propose a novel single-shot detector, called RFAlignNet. Our model is composed of two sequentially inter-connected layers, i.e, anchor refinement layer and receptive field alignment detection head. The former layer coarsely refines the locations and sizes of the pre-defined anchors to provide better anchor set to latter detection head, reducing the issue of sensitiveness of pre-defined anchors. The latter layer further exploits the refined anchors to align the receptive field of detection head onto the object-related region, avoiding the misalignment issue. Our results show that with the proposed cascaded inference mechanism, RFAlignNet outperforms the accuracy of state-of-the-art one-stage and two-stage detectors, while preserving the high efficiency. Our model runs in real-time with only 24.9M of parameters that are less than the one of lightweight SSD (26.5M).