Statistical adaptive modeling for kitchen waste detection in complex scenes.

Automatic detection of kitchen waste is of great significance, which provides support for its subsequent full quantitative consumption and harmless treatment. In addition, manual sorting of kitchen waste is inefficient and toxic waste is harmful to human health, making automatic detection of kitchen waste technology crucial. Automatic detection of kitchen waste in complex scenes faces the challenge of diversity of category outlines and uneven distribution. In this paper, we propose a detector based on statistical adaptive modeling(namely SA-Det) for the automatic detection of kitchen waste in complex scenes. Firstly, to solve the issue of diversity of category outlines, we propose a category statistics adaptive (CSA) module. The CSA module constructs dynamic thresholds for each instance to accurately assign positive and negative samples by fusing category statistics and instance shape information, thereby improving detection performance. Moreover, to solve the issue of uneven distribution, we propose a distribution adaptive (DA) module, which dynamically adjusts the loss weights by adaptively sensing the number of labels during training process. Extensive experiments on our constructed kitchen waste dataset (KWD) demonstrate that SA-Det consistently and significantly improves the performance of existing state-of-the-art methods (e.g., Rotated_RetinaNet (Lin et al., 2017) and RoI_Transformer (Ding et al., 2019)) by around 2% to 3.5%. • We propose a detector based on statistical adaptive modeling (namely SA-Det) for the automatic detection of kitchen waste in complex scenes. Extensive experiments on our constructed kitchen waste dataset (KWD) demonstrate that SA-Det consistently and significantly improves the performance of existing state-of-the-art methods by around 2% to 3.5%. • To solve the issue of diversity of category outlines, we design a category statistical adaptive (CSA) module by fusing the statistical characteristics of each category and instance shape information, which provides a dynamic threshold for each target instance to accurately allocate positive and negative samples and improve detection precision. • According to the distribution difference faced by kitchen waste in detection, we propose a distribution adaptive (DA) module, which dynamically allocates loss weights by adaptively sensing the number of targets during the training process, thereby effectively alleviating the issue of uneven distribution of kitchen waste. [ABSTRACT FROM AUTHOR]