Performing Weakly Supervised Retail Instance Segmentation via Region Normalization

Instance segmentation can clearly distinguish object instances from the pixel level, which is the foundation for applications in the retail industry. Supervised instance segmentation methods require pixel-level annotations for learning accurate object patterns, which is expensive and labor-intensive to obtain. In the case of retail industry, objects of interest vary frequently and only category-level labels are available, which motivates us to study the weakly supervised instance segmentation algorithm for the retail industry. Although weakly supervised instance segmentation algorithms have been extensively studied on standard benchmarks, e.g., VOC and COCO, directly employing these algorithms to the retail industry usually generates insufficient segmentation accuracy. We found a fundamental reason lies in the specific characteristics of the retail industry. Specifically, one challenge for the traditional benchmark dataset is dramatic appearance variations, but the retail industry only tackles limited object categories and each object possesses fixed appearances. To thoroughly explore characteristics of the retail industry, we propose the region normalization mechanism to thoroughly explore characteristics of the retail industry. It divides an image into regions and requires pixels from the same region to possess consistent features. The region normalization mechanism is a novel attempt to perform normalization within irregular regions, distinct from the existing normalization mechanism. Besides, as region division can influence the effectiveness of the region normalization operation, we explore feature information to dynamically divide an image into regions, under constraints that pixels belonging to the same region should show similar features and keep spatially adjacent. The proposed region normalization mechanism and region division strategy endow a simple but effective solution to weakly supervised instance segmentation for the retail industry. On a large-scale benchmark dataset MVTec D2S, the proposed method performs favorably over existing well-performed methods, and the effectiveness of the region normalization mechanism is demonstrated as well.