DeCo-DETR: Densely Packed Commodity Detection with Transformer
Linhui Dai, Hong Liu · 2023
Commodity object detection (COD) plays a significant role in the development of smart unmanned supermarkets. Due to the dense arrangement of items, COD is more challenging than classic object detection. Traditional object detection methods often require intricate hand-designed steps and predefined anchors. These steps increase the difficulty of detecting objects arranged densely. In this paper, we present a DEnsely-packed COmmodity DEtection TRansformer (DeCo-DETR) framework for commodity object detection. Firstly, we propose an adaptive positional prior generator module, which uses anchor box sizes to adjust cross-attention. This enables the cross-attention module to concentrate on local regions corresponding to the target objects. A density-map-guided ranker assignment strategy is introduced to prevent incorrect associations between ground truth boxes and predicted boxes. This strategy treats the object box as a Gaussian distribution and computes the Gaussian Wasserstein distance based on that distribution. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on the largest dense product dataset, SKU-110K.