Weakly Supervised Object Localization Based on Implicit Spatial Constraints

Weakly Supervised Object Localization (WSOL) tasks, as one of the most challenging tasks in the field of computer vision, aims to locate objects using only a small number of image-level labels, thus reducing annotation costs. The most popular paradigm in current weakly supervised object detection divides the WSOL task into two parts: class-agnostic object localization and object classification. After continuous optimization by subsequent scholars, models based on the Shallow Feature-aware Pseudo Label Supervised Object Localization (SPOL) paradigm have shown good performance. In this paper, we propose a spatial awareness attention module to construct implicit object spatial feature relationships in images, obtaining clear object boundaries as constraints, and then using a multi-dimensional convolutional attention mechanism to diffuse the target activation area. This method is designed to solve the phenomenon of class partial activation that existed in previous methods, because it can obtain clear object boundary information, which affects the activation of the overall target area. In addition, we use Gaussian mixture modeling for class-agnostic model segmentation to achieve precise object masks, which can overcome the negative impact of multiple objects and background noise on mask generation. Experiments verify that our model outperforms the baseline model on both the CUB-200–2011 and ImageNet-1K benchmarks, achieving 96.42% and 69.04% on the GT-known metric respectively (increases of 0.32% and 1.15%).